Re: [PATCH] virtio_ring: use smp_store_mb

On Thu, Dec 17, 2015 at 03:52:24PM +0000, Will Deacon wrote:

On Thu, Dec 17, 2015 at 04:09:17PM +0100, Peter Zijlstra wrote:

quoted

On Thu, Dec 17, 2015 at 04:34:57PM +0200, Michael S. Tsirkin wrote:

quoted

On Thu, Dec 17, 2015 at 03:02:12PM +0100, Peter Zijlstra wrote:

quoted

quoted

quoted

	commit 9e1a27ea42691429e31f158cce6fc61bc79bb2e9
	Author: Alexander Duyck [off-list ref]
	Date:   Mon Apr 13 21:03:49 2015 +0930

	    virtio_ring: Update weak barriers to use dma_wmb/rmb

That commit doesn't make any sense. dma_*mb() explicitly does _NOT_
cover the smp_*mb() part.

Again, look at the ARM definitions, the smp_*mb() primitives use the
inner coherence stuff, while the dma_*mb() primitives use the outer
coherent stuff.

Does outer coherent imply inner coherent?

quoted

the *mb() primitives cover both.

I do not think so, but lets add Will, he dreams this stuff.

Right, and I don't sleep well these days.

Anyway, the outer-shareable domain (osh) is a superset of the
inner-shareable domain (ish). The inner-shareable domain contains the
CPUs and any peripherals that you and I would call "cache coherent". The
outer-shareable domain extends this to cover a strange set of "less cache
coherent" devices, which we would just call "not cache coherent" for the
purposes of Linux. Normal, non-cacheable memory (i.e. the memory type we
use for non-coherent DMA buffers) is outer-shareable.

Since the barrier macros don't know if the device is coherent or not, we
use the stronger semantics of outer-shareable.

I've not been following the thread, but I reckon we could add dma_mb()
(as dmb(osh) on arm), then use that to build dma_load_acquire and
dma_store_release accessors. On arm64, we could probably use the
acquire/release instructions directly, since they inherit the shareability
domain of the address (which has the nice property of being inner-shareable
for coherent devices).

The massive pain with adding new accessors is defining the semantics.
memory-barriers.txt is already lacking for the CPU side, and we're
struggling to express the kind of transitivity guarantees we provide
today, let alone with new primitives :(

Will

Well virtio (might) have wanted dma_mb in the past.

But if are adding barrier stuff anyway, we really want
pv_ counterparts to smp_ that do the same on CONFIG_SMP
but don't change when CONFIG_SMP is disabled.

-- 
MST