[PATCH v7 3/8] arm64: introduce is_device_dma_coherent

[Stefano Stabellini]

On Fri, 7 Nov 2014, Catalin Marinas wrote:
> On Thu, Nov 06, 2014 at 12:22:13PM +0000, Stefano Stabellini wrote:
> > On Thu, 6 Nov 2014, Catalin Marinas wrote:
> > > On Wed, Nov 05, 2014 at 06:15:38PM +0000, Stefano Stabellini wrote:
> > > > On Wed, 5 Nov 2014, Catalin Marinas wrote:
> > > > > Note that if !is_device_dma_coherent(), it doesn't always mean that
> > > > > standard cache maintenance would be enough (but that's a Xen problem,
> > > > > not sure how to solve).
> > > > 
> > > > It is a thorny issue indeed.
> > > > Xen would need to know how to do non-standard cache maintenance
> > > > operations.
> > > 
> > > Is EL2 hyp or EL1 dom0 doing such maintenance? If the latter, you could
> > > just use the currently registered dma ops.
> > 
> > It is Xen, so El2 hypervisor.
> 
> So what is arch/arm/xen/mm32.c cache maintenance for? Doesn't it run in
> dom0?

This patch series changes that code path specifically. As a matter of
fact it removes mm32.c.
Before this patch series, cache maintenance in dom0 was being done with
XENFEAT_grant_map_identity (see explanation in previous email).
After this patch series, an hypercall is made when foreign pages are
involved, see patch 8/8, the if(!pfn_valid(pfn)) code path. Local pages
are still flushed in dom0.


> > > > Otherwise we would need to resurrect XENFEAT_grant_map_identity (that I
> > > > am reverting in this series) and be content with having CONFIG_XEN_DOM0
> > > > depend on CONFIG_ARM_LPAE.
> > > 
> > > So what does buy you? Is it just the hope that with LPAE you won't have
> > > weird system caches?
> > 
> > Let me explain a bit more and let's assume there is no SMMU.
> > 
> > The problem is not normal dma originating in dom0, currently mapped 1:1
> > (pseudo-physical == physical). The problem are dma operations that
> > involve foreign pages (pages belonging to other virtual machines)
> > currently mapped also in dom0 to do IO. It is common for the Xen PV
> > network and block frontends to grant access to one or more pages to the
> > network and block backends for IO. The backends, usually in dom0, map
> > the pages and perform the actual IO. Sometimes the IO is a dma operation
> > that involves one of the granted pages directly.
> > 
> > For these pages, the pseudo-physical address in dom0 is different from
> > the physical address. Dom0 keeps track of the pseudo-physical to
> > physical relationship (pfn_to_mfn in Xen terminology). The swiotlb-xen
> > driver makes sure to return the right mfn from map_page and map_sg.
> 
> For my understanding, xen_dma_map_page() is called from dom0 and it
> simply calls the __generic_dma_ops()->map_page(). The "page" argument
> here is the dom0 page and it gets translated to dom0 phys and then to a
> bus address via xen_phys_to_bus().
> 
> On arm64, __swiotlb_map_page() uses the page structure to get some
> pseudo device address which gets converted back to a dom0 virtual
> address. The latter is used for cache maintenance by VA in dom0. With
> PIPT-like caches, that's fine, you don't need the actual machine PA
> unless you have caches like PL310 (which are not compatible with
> virtualisation anyway and the latest ARMv8 ARM even makes a clear
> statement here).
> 
> (one potential problem here is the dma_capable() check in
> swiotlb_map_page() (non-xen) which uses the pseudo device address)
> 
> The interesting part is that xen_swiotlb_map_page() returns the real
> machine device address to dom0, which is different from what dom0 thinks
> of a device address (phys_to_dma(phys_to_page(page))). Does dom0 use
> such real/machine device address to program the device directly or there
> is another layer where you could do the translation?

No other layer. Dom0 uses the real/machine device address returned by
xen_phys_to_bus to program the device directly.


> > However some of the dma_ops give you only a dma_addr_t handle
> > (unmap_page and unmap_sg), that is the physical address of the page.
> 
> OK, I started to get it now, the __swiotlb_unmap_page() on arm64, if
> called from dom0, would get the machine device address and dma_to_phys()
> does not work.

Right!


> > Dom0 doesn't know how to find the pseudo-physical address corresponding
> > to it. Therefore it also doesn't know how to find the struct page for
> > it. This is not a trivial problem to solve as the same page can be
> > granted multiple times, therefore could have multiple pseudo-physical
> > addresses and multiple struct pages corresponding to one physical
> > address in dom0.
> 
> So what is mfn_to_pfn() and xen_bus_to_phys() used for?

They are not used on ARM. But swiotlb-xen is not ARM specific and
xen_bus_to_phys is still important on x86. It works differently there.


> Isn't mfn_to_pfn(pfn_to_mfn(x)) an identity function?

Yes, it should be and it is for local pages.
However it is not an identity function for foreign pages, because
mfn_to_pfn is broken and just returns pfn on ARM, while pfn_to_mfn works
correctly.


> > Fortunately these dma_ops don't actually need to do much. In fact they
> > only need to flush caches if the device is not dma-coherent. So the
> > current proposed solution is to issue an hypercall to ask Xen to do the
> > flushing, passing the physical address dom0 knows.
> 
> I really don't like the dma_cache_maint() duplication you added in
> arch/arm/xen/mm32.c just because you cannot call the host dma_ops when
> the page is not available.

I agree is bad, but this patch series is a big step forward removing the
duplication. In fact now that I think about it, when a page is local (not a
foreign page) I could just call the host dma_ops. mm.c:dma_cache_maint
would only be used for foreign pages (pages for which !pfn_valid(pfn)).


> You basically only need a VA that was mapped in dom0 when map_page() was
> called and is still mapped when page_unmap() is called. You can free the
> actual mapping after calling __generic_dma_ops()->unmap_page() in
> xen_dma_unmap_page() (in xen_unmap_single()).

That is true, but where would I store the mapping?

I don't want to introduce another data structure to keep physical to va
or physical to struct page mappings that I need to modify at every
map_page and unmap_page. It wouldn't even be a trivial data structure as
multiple dma operations involving the same mfn but different struct page
can happen simultaneously.

The performance hit of maintaining such a data structure would be
significant.

It would negatively impact any dma operations involving any devices,
including dma coherent ones. While this series only requires special
handling of dma operations involving non-coherent devices (resulting in
an hypercall being made).

In a way this series rewards hardware that makes life easier to software
developers :-)