[PATCH v7 3/8] arm64: introduce is_device_dma_coherent

[Catalin Marinas]

On Fri, Nov 07, 2014 at 05:35:41PM +0000, Stefano Stabellini wrote:
> On Fri, 7 Nov 2014, Stefano Stabellini wrote:
> > On Fri, 7 Nov 2014, Catalin Marinas wrote:
> > > What I would like to see is xen_dma_map_page() also using hyp calls for
> > > cache maintenance when !pfn_valid(), for symmetry with unmap. You would
> > > need another argument to xen_dma_map_page() though to pass the real
> > > device address or mfn (and on the map side you could simply check for
> > > page_to_pfn(page) != mfn). For such cases, Xen swiotlb already handles
> > > bouncing so you don't need dom0 swiotlb involved as well.
> > 
> > I can see that it would be nice to have map_page and unmap_page be
> > symmetrical. However actually doing the map_page flush with an hypercall
> > would slow things down. Hypercalls are slower than function calls. It is
> > faster to do the cache flushing in dom0 if possible. In the map_page
> > case we have the struct page so we can easily do it by calling the
> > native dma_ops function.
> > 
> > Maybe I could just add a comment to explain the reason for the asymmetry?
> 
> Ah, but the problem is that map_page could allocate a swiotlb buffer
> (actually it does on arm64) that without a corresponding unmap_page
> call, would end up being leaked, right?

Yes. You could hack dma_capable() to always return true for dom0
(because the pfn/dma address here doesn't have anything to do with the
real mfn) but that's more of a hack assuming a lot about the swiotlb
implementation.

> Oh well.. two hypercalls it is then :-/

That looks cleaner to me (though I haven't seen the code yet ;)).

Alternatively, you could keep a permanent page (per-cpu) in dom0 that
you ask the hypervisor to point at the mfn just for the unmap cache
maintenance (similarly to the kernel kmap_atomic used for cache
maintenance on high pages). This could be more expensive but, OTOH, you
only need the hyp call on the unmap path. The is_device_dma_coherent()
would still remain to avoid any unnecessary map/unmap calls but for
non-coherent devices you use whatever the SoC provides.

-- 
Catalin