On Thu, 6 Apr 2023 at 06:09, Paolo Bonzini <pbonz...@redhat.com> wrote: > > Replace with an explicit barrier and a comment. > > Signed-off-by: Paolo Bonzini <pbonz...@redhat.com> > --- > util/qemu-coroutine.c | 10 +++++++--- > 1 file changed, 7 insertions(+), 3 deletions(-) > > diff --git a/util/qemu-coroutine.c b/util/qemu-coroutine.c > index 849452369201..17a88f65053e 100644 > --- a/util/qemu-coroutine.c > +++ b/util/qemu-coroutine.c > @@ -127,9 +127,13 @@ void qemu_aio_coroutine_enter(AioContext *ctx, Coroutine > *co) > Coroutine *to = QSIMPLEQ_FIRST(&pending); > CoroutineAction ret; > > - /* Cannot rely on the read barrier for to in aio_co_wake(), as there > are > - * callers outside of aio_co_wake() */ > - const char *scheduled = qatomic_mb_read(&to->scheduled); > + /* > + * Read to before to->scheduled; pairs with qatomic_cmpxchg in > + * qemu_co_sleep(), aio_co_schedule() etc. > + */ > + smp_read_barrier_depends();
I'm not a fan of nuanced memory ordering primitives. I don't understand or remember all the primitives available in docs/devel/atomics.rst and especially not how they interact with each other. Does smp_read_barrier_depends() make sense for QEMU? Does QEMU support Alpha host CPUs? Stefan
