Re: [PATCH v2 1/1] mm: fix the theoretical compound_lock() vs prep_new_page() race

[Oleg Nesterov <oleg@redhat.com>]

On 01/09, Andrea Arcangeli wrote:
>
> >
> > But we probably need barrier() in between, we can't use ACCESS_ONCE().
>
> After get_page_unless_zero I don't think there's any need of
> barrier(). barrier() should have been implicit in __atomic_add_unless
> in fact it should be a full smp_mb() equivalent too. Memory is always
> clobbered there and the asm is volatile.

Yes, yes,

> My wondering was only about the runtime (not compiler) barrier after
> running PageTail and before compound_lock,

Yes, this is what I meant.

Except I really meant the compiler barrier, although I do not think it
is actually needed, test_and_set_bit() implies mb().

> because bit_spin_lock has
> only acquire semantics so in absence of the branch that bails out the
> lock, the spinlock could run before PageTail. If the branch is good
> enough guarantee for all archs it's good and cheap solution.

The recent "[PATCH v6 tip/core/locking 3/8] Documentation/memory-barriers.txt:
Prohibit speculative writes" from Paul says:

	No SMP architecture currently supporting Linux allows speculative writes,

	...

	+ACCESS_ONCE(), which preserves the ordering between
	+the load from variable 'a' and the store to variable 'b':
	+
	+       q = ACCESS_ONCE(a);
	+       if (q) {
	+               ACCESS_ONCE(b) = p;
	+               do_something();
	+       }


We can't use ACCESS_ONCE(), but I think that

		if (PageTail(page)) {
			barrier();
			compound_lock(page_head);
		}

should obviously work (even if compound_lock() didn't imply mb).

Oleg.