Nouveau - Feb 2021 - [PATCH RFC v1 5/6] xen-swiotlb: convert variables to arrays

On Thu, Feb 04, 2021 at 09:40:23AM +0100, Christoph Hellwig
wrote:
> So one thing that has been on my mind for a while:  I'd really like
> to kill the separate dma ops in Xen swiotlb.  If we compare xen-swiotlb
> to swiotlb the main difference seems to be:
> 
>  - additional reasons to bounce I/O vs the plain DMA capable
>  - the possibility to do a hypercall on arm/arm64
>  - an extra translation layer before doing the phys_to_dma and vice
>    versa
>  - an special memory allocator
> 
> I wonder if inbetween a few jump labels or other no overhead enablement
> options and possibly better use of the dma_range_map we could kill
> off most of swiotlb-xen instead of maintaining all this code duplication?
So I looked at this a bit more.

For x86 with XENFEAT_auto_translated_physmap (how common is that?)
pfn_to_gfn is a nop, so plain phys_to_dma/dma_to_phys do work as-is.

xen_arch_need_swiotlb always returns true for x86, and
range_straddles_page_boundary should never be true for the
XENFEAT_auto_translated_physmap case.

So as far as I can tell the mapping fast path for the
XENFEAT_auto_translated_physmap can be trivially reused from swiotlb.

That leaves us with the next more complicated case, x86 or fully cache
coherent arm{,64} without XENFEAT_auto_translated_physmap.  In that case
we need to patch in a phys_to_dma/dma_to_phys that performs the MFN
lookup, which could be done using alternatives or jump labels.
I think if that is done right we should also be able to let that cover
the foreign pages in is_xen_swiotlb_buffer/is_swiotlb_buffer, but
in that worst case that would need another alternative / jump label.

For non-coherent arm{,64} we'd also need to use alternatives or jump
labels to for the cache maintainance ops, but that isn't a hard problem
either.

On Sun, Feb 07, 2021 at 04:56:01PM +0100, Christoph Hellwig
wrote:
> On Thu, Feb 04, 2021 at 09:40:23AM +0100, Christoph Hellwig wrote:
> > So one thing that has been on my mind for a while:  I'd really
like
> > to kill the separate dma ops in Xen swiotlb.  If we compare
xen-swiotlb
> > to swiotlb the main difference seems to be:
> > 
> >  - additional reasons to bounce I/O vs the plain DMA capable
> >  - the possibility to do a hypercall on arm/arm64
> >  - an extra translation layer before doing the phys_to_dma and vice
> >    versa
> >  - an special memory allocator
> > 
> > I wonder if inbetween a few jump labels or other no overhead
enablement
> > options and possibly better use of the dma_range_map we could kill
> > off most of swiotlb-xen instead of maintaining all this code
duplication?
> 
> So I looked at this a bit more.
> 
> For x86 with XENFEAT_auto_translated_physmap (how common is that?)
Juergen, Boris please correct me if I am wrong, but that
XENFEAT_auto_translated_physmap
only works for PVH guests?
> pfn_to_gfn is a nop, so plain phys_to_dma/dma_to_phys do work as-is.
> 
> xen_arch_need_swiotlb always returns true for x86, and
> range_straddles_page_boundary should never be true for the
> XENFEAT_auto_translated_physmap case.
Correct. The kernel should have no clue of what the real MFNs are
for PFNs.
> 
> So as far as I can tell the mapping fast path for the
> XENFEAT_auto_translated_physmap can be trivially reused from swiotlb.
> 
> That leaves us with the next more complicated case, x86 or fully cache
> coherent arm{,64} without XENFEAT_auto_translated_physmap.  In that case
> we need to patch in a phys_to_dma/dma_to_phys that performs the MFN
> lookup, which could be done using alternatives or jump labels.
> I think if that is done right we should also be able to let that cover
> the foreign pages in is_xen_swiotlb_buffer/is_swiotlb_buffer, but
> in that worst case that would need another alternative / jump label.
> 
> For non-coherent arm{,64} we'd also need to use alternatives or jump
> labels to for the cache maintainance ops, but that isn't a hard problem
> either.
> 
>