Re: rcu_read_lock lost its compiler barrier

[Linus Torvalds]

On Mon, Jun 3, 2019 at 8:26 AM David Laight <David.Laight@xxxxxxxxxx> wrote:
>
> From: Paul E. McKenney
>
> > We do
> > occasionally use READ_ONCE() to prevent load-fusing optimizations that
> > would otherwise cause the compiler to turn while-loops into if-statements
> > guarding infinite loops.
>
> In that case the variable ought to be volatile...

No.

We do not use volatile on variables.

The C people got the semantics wrong, probably because 'volatile' was
historically for IO, not for access atomicity without locking.

It's not the memory location that is volatile, it is really the
_access_ that is volatile.

The same memory location might be completely stable in other access
situations (ie when done under a lock).

In other words, we should *never* use volatile in the kernel. It's
fundamentally mis-designed for modern use.

(Of course, we then can use volatile in a cast in code, which drives
some compiler people crazy, but that's because said compiler people
don't care about reality, they care about some paperwork).

                      Linus