Re: [PATCH] mmc: sdhci-xenon: Fix 2G limitation on AC5 SoC

[Robin Murphy]

On 2022-08-16 21:51, Vadym Kochan wrote:
[...]
> > The one thing to watch out for is that SWIOTLB doesn't necessarily interact
> > very well with DMA offsets. Given the intent of
> > of_dma_get_max_cpu_address(), I think it ought to work out OK now for
> > current kernels on DT systems if everything is described correctly, but
> > otherwise it's likely that you end up with ZONE_DMA either being empty or
> > containing all memory, so the SWIOTLB buffer ends up being allocated
> > anywhere such that it might not actually work as expected.
> >
> > Robin.
>
> Hi Robin,
>
> Thank you for the reply.
>
> My understanding is that swiotlb is allocated (in case of arm64)
> in the following cases:
>
>     #1 when it is forced from the kernel cmdline
>
>     #2 when max_pfn is greater than arm64_dma_phys_limit (and this is used
>        as the end from which to allocate the swiotlb pool in the
>        top-botom direction via memblock API).
>
>     #3 using restricted dma-pool
>
> Of course option #3 works fine because swiotlb is kind of forced to use
> particulary this range of memory.
>
> Both options #1 & #2 causes to use full memory mask even if to specify
> dma-ranges in the DT:
>
>      dma-ranges = <0x0 0x0 0x2 0x0 0x0 0x80000000>;
>
> or if to specify the opposite:
>
>      dma-ranges = <0x2 0x0 0x0 0x0 0x0 0x80000000>;
>
>      just to make it lower than U32 to pass
>
>          zone_dma_bits = min3(32U, dt_zone_dma_bits, acpi_zone_dma_bits)
>
>      condition, but then it will be re-set in max_zone_phys() by:
>
>       
> 	if (phys_start > U32_MAX)
> 		zone_mask = PHYS_ADDR_MAX;
> 	else if (phys_start > zone_mask)
> 		zone_mask = U32_MAX;

Ah, indeed I missed that, sorry. It seems that that change to stop 
assuming an offset kind of crossed over with the introduction of 
*_dma_get_max_cpu_address(), but now that that firmware property parsing 
*is* implemented, in principle it should be equally possible to evaluate 
the actual offsets as well, and decide whether an offset ZONE_DMA is 
appropriate or not. Either way, this is definitely the area which needs 
work if we want to to able to support topologies like this properly.

> So, currently I dont see how to pin swiotlb (I see it as a main problem) to some specific range of physical
> memory (particulary to the first 2G of RAM).

Indeed, if ZONE_DMA and/or ZONE_DMA32 can't be set appropriately, then 
there's no way to guarantee correct allocation of any DMA buffers, short 
of hacking it with explicitly placed reserved-memory carveouts.

Thanks,
Robin.