Re: [PATCH] fs: add memory barrier in __fget_light()

[Mark Rutland]

On Mon, Oct 31, 2022 at 10:37:01AM -0700, Linus Torvalds wrote:
> On Mon, Oct 31, 2022 at 10:13 AM Jann Horn <jannh@xxxxxxxxxx> wrote:
> >
> > If this is too expensive on platforms like arm64, I guess the more
> > performant alternative would be to add another flags field that tracks
> > whether the fs_struct was ever shared and check that instead of the
> > reference count in __fget_light().
> 
> No, the problem is that you should never use the "smp_*mb()" horrors
> for any new code.
> 
> All the "smp_*mb()" things really are broken. Please consider them
> legacy garbage. It was how people though about SMP memory ordering in
> the bad old days.
> 
> So get with the 21st century, and instead replace the "atomic_read()"
> with a "smp_load_acquire()".

Minor nit: atomic{,64,_long}_{read_acquire,set_release}() exist to be used
directly on atomics and should d.t.r.t. on all architectures (e.g. where 64-bit
atomics on 32-bit platforms have extra requirements).

So this instance can be:

  ...
  if (atomic_read_acquire(&files->count) == 1) {
  ...

Mark.