Re: [RFC] Disable lockref on arm64

[Ard Biesheuvel]

On Fri, 14 Jun 2019 at 11:58, Will Deacon <will.deacon@xxxxxxx> wrote:
>
> [+Kees]
>
> On Fri, Jun 14, 2019 at 07:09:26AM +0000, Jayachandran Chandrasekharan Nair wrote:
> > On Wed, Jun 12, 2019 at 10:31:53AM +0100, Will Deacon wrote:
> > > On Wed, Jun 12, 2019 at 04:10:20AM +0000, Jayachandran Chandrasekharan Nair wrote:
> > > > Now that the lockref change is mainline, I think we need to take another
> > > > look at this patch.
> > >
> > > Before we get too involved with this, I really don't want to start a trend of
> > > "let's try to rewrite all code using cmpxchg() in Linux because of TX2".
> >
> > x86 added a arch-specific fast refcount implementation - and the commit
> > specifically notes that it is faster than cmpxchg based code[1].
> >
> > There seems to be an ongoing effort to move over more and more subsystems
> > from atomic_t to refcount_t(e.g.[2]), specifically because refcount_t on
> > x86 is fast enough and you get some error checking atomic_t that does not
> > have.
>
> Correct, but there are also some cases that are only caught by
> REFCOUNT_FULL.
>

Yes, but do note that my arm64 implementation catches
increment-from-zero as well.

> > > At some point, the hardware needs to play ball. However...
> >
> > Even on a totally baller CPU, REFCOUNT_FULL is going to be slow :)
> > On TX2, this specific benchmark just highlights the issue, but the
> > difference is significant even on x86 (as noted above).
>
> My point was more general than that. If you want scalable concurrent code,
> then you end up having to move away from the serialisation introduced by
> locking. The main trick in the toolbox is cmpxchg() so, in the absence of
> a zoo of special-purpose atomic instructions, it really needs to do better
> than serialising.
>
> > > I was hoping we could use LDMIN/LDMAX to maintain the semantics of
> > > REFCOUNT_FULL, but now that I think about it I can't see how we could keep
> > > the arithmetic atomic in that case. Hmm.
> >
> > Do you think Ard's patch needs changes before it can be considered? I
> > can take a look at that.
>
> I would like to see how it performs if we keep the checking inline, yes.
> I suspect Ard could spin this in short order.
>

Moving the post checks before the stores you mean? That shouldn't be
too difficult, I suppose, but it will certainly cost performance.

> > > Whatever we do, I prefer to keep REFCOUNT_FULL the default option for arm64,
> > > so if we can't keep the semantics when we remove the cmpxchg, you'll need to
> > > opt into this at config time.
> >
> > Only arm64 and arm selects REFCOUNT_FULL in the default config. So please
> > reconsider this! This is going to slow down arm64 vs. other archs and it
> > will become worse when more code adopts refcount_t.
>
> Maybe, but faced with the choice between your micro-benchmark results and
> security-by-default for people using the arm64 Linux kernel, I really think
> that's a no-brainer. I'm well aware that not everybody agrees with me on
> that.

I think the question whether the benchmark is valid is justified, but
otoh, we are obsessed with hackbench which is not that representative
of a real workload either. It would be better to discuss these changes
in the context of known real-world use cases where refcounts are a
true bottleneck.

Also, I'd like to have Kees's view on the gap between REFCOUNT_FULL
and the fast version on arm64. I'm not convinced the cases we are not
covering are such a big deal.