[RFC] arch: Introduce new TSO memory barrier smp_tmb()

[Will Deacon]

On Sun, Nov 03, 2013 at 11:34:00PM +0000, Linus Torvalds wrote:
> So it would *kind* of act like a "smp_wmb() + smp_rmb()", but the
> problem is that a "smp_rmb()" doesn't really "attach" to the preceding
> write.

Agreed.

> This is analogous to a "acquire" operation: you cannot make an
> "acquire" barrier, because it's not a barrier *between* two ops, it's
> associated with one particular op.
> 
> So what I *think* you actually really really want is a "store with
> release consistency, followed by a write barrier".

How does that order reads against reads? (Paul mentioned this as a
requirement). I not clear about the use case for this, so perhaps there is a
dependency that I'm not aware of.

> In TSO, afaik all stores have release consistency, and all writes are
> ordered, which is why this is a no-op in TSO. And x86 also has that
> "all stores have release consistency, and all writes are ordered"
> model, even if TSO doesn't really describe the x86 model.
> 
> But on ARM64, for example, I think you'd really want the store itself
> to be done with "stlr" (store with release), and then follow up with a
> "dsb st" after that.

So a dsb is pretty heavyweight here (it prevents execution of *any* further
instructions until all preceeding stores have completed, as well as
ensuring completion of any ongoing cache flushes). In conjunction with the
store-release, that's going to hold everything up until the store-release
(and therefore any preceeding memory accesses) have completed. Granted, I
think that gives Paul his read/read ordering, but it's a lot heavier than
what's required.

> And notice how that requires you to mark the store itself. There is no
> actual barrier *after* the store that does the optimized model.
> 
> Of course, it's entirely possible that it's not worth worrying about
> this on ARM64, and that just doing it as a "normal store followed by a
> full memory barrier" is good enough. But at least in *theory* a
> microarchitecture might make it much cheaper to do a "store with
> release consistency" followed by "write barrier".

I agree with the sentiment but, given that this stuff is so heavily
microarchitecture-dependent (and not simple to probe), a simple dmb ish
might be the best option after all. That's especially true if the
microarchitecture decided to ignore the barrier options and treat everything
as `all accesses, full system' in order to keep the hardware design simple.

Will