Re: [RFC 2/3] iommu/samsung: Introduce Exynos sysmmu-v8 driver

[Robin Murphy]

On 2022-06-21 20:57, Sam Protsenko wrote:
> Hi Marek,
>
> On Fri, 21 Jan 2022 at 14:31, Marek Szyprowski <m.szyprowski@xxxxxxxxxxx> wrote:
>
> [snip]
>
> > Well, for starting point the existing exynos-iommu driver is really
> > enough. I've played a bit with newer Exyos SoCs some time ago. If I
> > remember right, if you limit the iommu functionality to the essential
> > things like mapping pages to IO-virtual space, the hardware differences
> > between SYSMMU v5 (already supported by the exynos-iommu driver) and v7
> > are just a matter of changing a one register during the initialization
> > and different bits the page fault reason decoding. You must of course
> > rely on the DMA-mapping framework and its implementation based on
> > mainline DMA-IOMMU helpers. All the code for custom iommu group(s)
> > handling or extended fault management are not needed for the initial
> > version.
>
> Thanks for the advice! Just implemented some testing driver, which
> uses "Emulated Translation" registers available on SysMMU v7. That's
> one way to verify the IOMMU driver with no actual users of it. It
> works fine with vendor SysMMU driver I ported to mainline earlier, and
> now I'm trying to use it with exynos-sysmmu driver (existing
> upstream). If you're curious -- I can share the testing driver
> somewhere on GitHub.
>
> I believe the register you mentioned is PT_BASE one, so I used
> REG_V7_FLPT_BASE_VM = 0x800C instead of REG_V5_PT_BASE_PFN. But I
> didn't manage to get that far, unfortunately, as
> exynos_iommu_domain_alloc() function fails in my case, with BUG_ON()
> at this line:
>
>      /* For mapping page table entries we rely on dma == phys */
>      BUG_ON(handle != virt_to_phys(domain->pgtable));
>
> One possible explanation for this BUG is that "dma-ranges" property is
> not provided in DTS (which seems to be the case right now for all
> users of "samsung,exynos-sysmmu" driver). Because of that the SWIOTLB
> is used for dma_map_single() call (in exynos_iommu_domain_alloc()
> function), which in turn leads to that BUG. At least that's what
> happens in my case. The call chain looks like this:
>
>      exynos_iommu_domain_alloc()
>          v
>      dma_map_single()
>          v
>      dma_map_single_attrs()
>          v
>      dma_map_page_attrs()
>          v
>      dma_direct_map_page()  // dma_capable() == false
>          v
>      swiotlb_map()
>          v
>      swiotlb_tbl_map_single()
>
> And the last call of course always returns the address different than
> the address for allocated pgtable. E.g. in my case I see this:
>
>      handle = 0x00000000fbfff000
>      virt_to_phys(domain->pgtable) = 0x0000000880d0c000
>
> Do you know what might be the reason for that? I just wonder how the
> SysMMU driver work for all existing Exynos platforms right now. I feel
> I might be missing something, like some DMA option should be enabled
> so that SWIOTLB is not used, or something like that. Please let me
> know if you have any idea on possible cause. The vendor's SysMMU
> driver is kinda different in that regard, as it doesn't use
> dma_map_single(), so I don't see such issue there.

If this SysMMU version is capable of addressing more than 32 bits, then 
exynos_iommu_probe_device() should set its DMA masks appropriately.

(as a side note since I looked, the use of PAGE_SIZE/PAGE_SHIFT in the 
driver looks wrong, since I can't imagine that the hardware knows 
whether Linux is using 4KB, 16KB or 64KB and adjusts itself accordingly...)

Robin.