Re: [RFC PATCH 2/3] iommu/dma-iommu: Support iovad->granule > PAGE_SIZE
From: Sven Peter
Date: Sat Aug 07 2021 - 07:48:12 EST
On Fri, Aug 6, 2021, at 20:04, Robin Murphy wrote:
> On 2021-08-06 16:55, Sven Peter via iommu wrote:
> > @@ -1006,6 +1019,31 @@ static int iommu_dma_map_sg(struct device *dev, struct scatterlist *sg,
> > if (dev_is_untrusted(dev))
> > return iommu_dma_map_sg_swiotlb(dev, sg, nents, dir, attrs);
> >
> > + /*
> > + * If the IOMMU pagesize is larger than the CPU pagesize we will
> > + * very likely run into sgs with a physical address that is not aligned
> > + * to an IOMMU page boundary. Fall back to just mapping every entry
> > + * independently with __iommu_dma_map then.
>
> Scatterlist segments often don't have nicely aligned ends, which is why
> we already align things to IOVA granules in main loop here. I think in
> principle we'd just need to move the non-IOVA-aligned part of the
> address from sg->page to sg->offset in the temporary transformation for
> the rest of the assumptions to hold. I don't blame you for being timid
> about touching that, though - it took me 3 tries to get right when I
> first wrote it...
>
I've spent some time with that code now and I think we cannot use it
but have to fall back to iommu_dma_map_sg_swiotlb (even though that swiotlb
part is a lie then):
When we have sg_phys(s) = 0x802e65000 with s->offset = 0 the paddr
is aligned to PAGE_SIZE but has an offset of 0x1000 from something
the IOMMU can map.
Now this would result in s->offset = -0x1000 which is already weird
enough.
Offset is unsigned (and 32bit) so this will actually look like
s->offset = 0xfffff000 then, which isn't much better.
And then sg_phys(s) = 0x902e64000 (instead of 0x802e64000) and
we'll map some random memory in iommu_map_sg_atomic and a little bit later
everything explodes.
Now I could probably adjust the phys addr backwards and make sure offset is
always positive (and possibly larger than PAGE_SIZE) and later restore it
in __finalise_sg then but I feel like that's pushing this a little bit too far.
Sven