On Wed, Jul 11, 2018 at 06:05:51PM +0800, Jiaxun Yang wrote:
> On 2018-7-10 Tue at 20:17:27，Peter Zijlstra Wrote：
> Hi Peter
> Since Huacai unable to send email via client, I'm going to reply for him
> > Sure.. we all got that far. And no, this isn't the _real_ problem. This
> > is a manifestation of the problem.
> > The problem is that your SFB is broken (per the Linux requirements). We
> > require that stores will become visible. That is, they must not
> > indefinitely (for whatever reason) stay in the store buffer.
> > > I don't think this is a hardware bug, in design, SFB will flushed to
> > > L1 cache in three cases:
> > >
> > > 1, data in SFB is full (be a complete cache line);
> > > 2, there is a subsequent read access in the same cache line;
> > > 3, a 'sync' instruction is executed.
> > And I think this _is_ a hardware bug. You just designed the bug instead
> > of it being by accident.
> Yes, we understood that this hardware feature is not supported by LKML,
> so it should be a hardware bug for LKML.
> > It doesn't happen an _any_ other architecture except that dodgy
> > ARM11MPCore part. Linux hard relies on stores to become available
> > _eventually_.
> > Still, even with the rules above, the best work-around is still the very
> > same cpu_relax() hack.
> As you say, SFB makes Loongson not fully SMP-coherent.
> However, modify cpu_relax can solve the current problem,
> but not so straight forward. On the other hand, providing a Loongson-specific
> WRITE_ONCE looks more reasonable, because it the eliminate the "non-cohrency".
> So we can solve the bug from the root.
Curious, but why is it not straight-forward to hack cpu_relax()? If you try
to hack WRITE_ONCE, you also need to hack atomic_set, atomic64_set and all
the places that should be using WRITE_ONCE but aren't ;~)