[PATCH RFC] arm64: DMA zone above 4GB

[Catalin Marinas]

On Sun, Nov 12, 2023 at 09:25:46AM +0200, Baruch Siach wrote:
> On Thu, Nov 09 2023, Catalin Marinas wrote:
> > On Wed, Nov 08, 2023 at 07:30:22PM +0200, Baruch Siach wrote:
> >> Consider a bus with this 'dma-ranges' property:
> >> 
> >>   #address-cells = <2>;
> >>   #size-cells = <2>;
> >>   dma-ranges = <0x00000000 0xc0000000 0x00000008 0x00000000 0x0 0x40000000>;
> >> 
> >> Devices under this bus can see 1GB of DMA range between 3GB-4GB. This
> >> range is mapped to CPU memory at 32GB-33GB.
> >
> > Is this on real hardware or just theoretical for now (the rest of your
> > email implies it's real)? Normally I'd expected the first GB (or first
> > two) of RAM from 32G to be aliased to the lower 32-bit range for the CPU
> > view as well, not just for devices. You'd then get a ZONE_DMA without
> > having to play with DMA offsets.
> 
> This hardware is currently in fabrication past tapeout. Software tests
> are running on FPGA models and software simulators.
> 
> CPU view of the 3GB-4GB range is not linear with DMA addresses. That is,
> for offset N where 0 <= N <= 1GB, the CPU address 3GB+N does not map to
> the same physical location of DMA address 3GB+N. Hardware engineers are
> not sure this is fixable. So as is often the case we look at software to
> save us. After all, from hardware perspective this design "works".

We had a similar setup before with the AMD Seattle platform and the
random assumption that that the first 4GB of the start of DRAM are used
for ZONE_DMA32. We reverted that since.

> >> Current zone_sizes_init() code considers 'dma-ranges' only when it maps
> >> to RAM under 4GB, because zone_dma_bits is limited to 32. In this case
> >> 'dma-ranges' is ignored in practice, since DMA/DMA32 zones are both
> >> assumed to be located under 4GB. The result is that the stmmac driver
> >> DMA buffers allocation GFP_DMA32 flag has no effect. As a result DMA
> >> buffer allocations fail.
> >> 
> >> The patch below is a crude workaround hack. It makes the  DMA zone
> >> cover the 1GB memory area that is visible to stmmac DMA as follows:
> >> 
> >> [    0.000000] Zone ranges:
> >> [    0.000000]   DMA      [mem 0x0000000800000000-0x000000083fffffff]
> >> [    0.000000]   DMA32    empty
> >> [    0.000000]   Normal   [mem 0x0000000840000000-0x0000000bffffffff]
> >> ...
> >> [    0.000000] software IO TLB: mapped [mem 0x000000083bfff000-0x000000083ffff000] (64MB)
> >> 
> >> With this hack the stmmac driver works on my platform with no
> >> modification.
> >> 
> >> Clearly this can't be the right solutions. zone_dma_bits is now wrong for
> >> one. It probably breaks other code as well.
> >
> > zone_dma_bits ends up as 36 if I counted correctly. So DMA_BIT_MASK(36)
> > is 0xf_ffff_ffff and the phys_limit for your device is below this mask,
> > so dma_direct_optimal_gfp_mask() does end up setting GFP_DMA. However,
> > looking at how it sets GFP_DMA32, it is obvious that the code is not set
> > up for such configurations. I'm also not a big fan of zone_dma_bits
> > describing a mask that goes well above what the device can access.
> >
> > A workaround would be for zone_dma_bits to become a *_limit and sort out
> > all places where we compare masks with masks derived from zone_dma_bits
> > (e.g. cma_in_zone(), dma_direct_supported()).
> 
> I was also thinking along these lines. I wasn't sure I see the entire
> picture, so I hesitated to suggest a patch. Specifically, the assumption
> that DMA range limits are power of 2 looks deeply ingrained in the
> code. Another assumption is that DMA32 zone is in low 4GB range.

I think we can safely assume that the DMA range (difference between the
min and max addresses) is a power of two. Once the DMA/CPU offsets are added,
this assumption no longer holds. The kernel assumes that the bits in the
DMA range as seen by the device would match the CPU addresses. More
below.

> I can work on an RFC implementation of this approach.
> 
> Petr suggested a more radical solution of per bus DMA constraints to
> replace DMA/DMA32 zones. As Petr acknowledged, this does not look like a
> near future solution.

Petr's suggestion is more generic indeed but I reckon it will take a
long time to implement and upstream.

> > Alternatively, take the DMA offset into account when comparing the
> > physical address corresponding to zone_dma_bits and keep zone_dma_bits
> > small (phys offset subtracted, so in your case it would be 30 rather
> > than 36).
> 
> I am not following here. Care to elaborate?

In your case, zone_dma_bits would be 30 covering 1GB. When doing all
those comparison in the generic code, make sure that (1 << zone_dma_bits)
value is used relative to the CPU offset or the start of DRAM rather
than an absolute limit. Similarly for ZONE_DMA32, it wouldn't be the
32-bit of the CPU address but rather the 32-bit from the start of the
RAM, on the assumption that the hardware lifts the device access into
the correct physical address.

So, I think it's doable keeping zone_dma_bits as 'bits' rather than
limit, it just needs some updates throughout the generic code. Now, the
DMA API maintainers may have a different opinion. Robin was away on
holiday last week so I couldn't check with him.

-- 
Catalin