From: Peter Zijlstra
> Sent: 08 May 2023 09:39
>
> On Sun, May 07, 2023 at 04:04:23PM +0200, Jonas Oberhauser wrote:
> >
> > Am 4/25/2023 um 9:58 PM schrieb Linus Torvalds:
> > > Yes, I think Mark is right. It's not that 'old' might be wrong - that
> > > doesn't matter because cmpxchg will work it out - it's just that 'new'
> > > might not be consistent with the old value we then use.
> >
> > In the general pattern, besides the potential issue raised by Mark, tearing
> > may also be an issue (longer example inspired by a case we met at the end of
> > the mail) where 'old' being wrong matters.
>
> There is yet another pattern where it actually matters:
>
> old = READ_ONCE(*ptr);
> do {
> if (cond(old))
> return false;
>
> new = func(old);
> } while (!try_cmpxchg(ptr, &old, new));
>
> return true;
>
> In this case we rely on old being 'coherent'. The more obvious case is
> where it returns old (also not uncommon), but even if it just checks a
> (multi-bit) condition on old you don't want tearing.
It isn't as though READ_ONCE() is expensive.
For kernel/device driver code, while CSE is useful, you pretty
much always want the compiler to always do loads into local
variables.
It is rather a shame there isn't a compiler option that
avoids these unusual any annoying operations.
Since the current 'rules' seem to require READ_ONCE() and
WRITE_ONCE() be used as pairs, why not make the data 'volatile'?
That ought to be the same as using volatile casts on all accesses.
David
-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)
