On Sat, Apr 17, 2021 at 04:51:58PM +0200, Paolo Bonzini wrote:

On 16/04/21 09:09, Peter Zijlstra wrote:

quoted

Well, the obvious example would be seqlocks. C11 can't do them

Sure it can.  C11 requires annotating with (the equivalent of) READ_ONCE all
reads of seqlock-protected fields, but the memory model supports seqlocks
just fine.

How does that help?

IIRC there's two problems, one on each side the lock. On the write side
we have:

	seq++;
	smp_wmb();
	X = r;
	Y = r;
	smp_wmb();
	seq++;

Which C11 simply cannot do right because it does't have wmb. You end up
having to use seq_cst for the first wmb or make both X and Y (on top of
the last seq) a store-release, both options are sub-optimal.

On the read side you get:

	do {
	  s = seq;
	  smp_rmb();
	  r1 = X;
	  r2 = Y;
	  smp_rmb();
	} while ((s&1) || seq != s);

And then you get into trouble the last barrier, so the first seq load
can be load-acquire, after which the loads of X, Y come after, but you
need then to happen before the second seq load, for which you then need
seq_cst, or make X and Y load-acquire. Again, not optimal.

I have also seen *many* broken variants of it on the web. Some work on
x86 but are totally broken when you build them on LL/SC ARM64.