On Thu, Nov 16, 2017 at 09:58:04AM +0100, Peter Zijlstra wrote:
> On Wed, Nov 15, 2017 at 10:01:11PM +0100, Andrea Parri wrote:
>
> > > And in specific things like:
> > >
> > > 135e8c9250dd5
> > > ecf7d01c229d1
> > >
> > > which use the release of rq->lock paired with the next acquire of the
> > > same rq->lock to match with an smp_rmb().
> >
> > Those cycles are currently forbidden by LKMM _when_ you consider the
> > smp_mb__after_spinlock() from schedule(). See rfi-rel-acq-is-not-mb
> > from my previous email and Alan's remarks about cumul-fence.
>
> I'm not sure I get your point; and you all seem to forget I do not in
> fact speak the ordering lingo. So I have no idea what
> rfi-blah-blah or cumul-fence mean.
I expand on my comment. Consider the following test:
C T1
{}
P0(int *x, int *y, spinlock_t *s)
{
spin_lock(s);
WRITE_ONCE(*x, 1);
spin_unlock(s);
spin_lock(s);
WRITE_ONCE(*y, 1);
spin_unlock(s);
}
P1(int *x, int *y)
{
int r0;
int r1;
r0 = READ_ONCE(*y);
smp_rmb();
r1 = READ_ONCE(*x);
}
exists (1:r0=1 /\ 1:r1=0)
According to LKMM, the store to x happens before the store to y but there
is no guarantee that the former store propagate (to P1) before the latter
(which is what we need to forbid that state). As a result, that state in
the "exists" clause is _allowed_ by LKMM.
The LKMM encodes happens-before (or execution) ordering with a relation
named "hb", while it encodes "propagation ordering" with "cumul-fence".
Andrea
>
> I know rel-acq isn't smp_mb() and I don't think any of the above patches
> need it to be. They just need it do be a local ordering, no?
>
> Even without smp_mb__after_spinlock() we get that:
>
> spin_lock(&x)
> x = 1
> spin_unlock(&x)
> spin_lock(&x)
> y = 1
> spin_unlock(&x)
>
> guarantees that x happens-before y, right?
>
> And that should be sufficient to then order something else against, like
> for example:
>
> r2 = y
> smp_rmb()
> r1 = x
>
> no?
>
>
From 1584215047007658452@xxx Thu Nov 16 09:40:01 +0000 2017
X-GM-THRID: 1582046402032606032
X-Gmail-Labels: Inbox,Category Forums,HistoricalUnread