On Fri, Jun 27, 2025, at 15:19, Adrian Hunter wrote:

On 27/06/2025 13:22, yangzh0906@thundersoft.com wrote:

quoted

Dear Mr. Hunter,

Our platform supports 64-bit physical addressing, but the eMMC controller's SRAM-based DMA engine is constrained to a 32-bit address space. 
When using the standard SDHCI interface, which allocates DDR-based DMA buffers with 64-bit addresses, thedma_map_single() operation fails 
because the DMA engine cannot handle addresses beyond 32 bits.

dma_map_single() should always succeed on arm64 even if the buffer
is outside of the dma mask: On most modern SoCs there is an SMMU/IOMMU
that implements the actual mapping, and in the absence of that
there is a fallback to SWIOTLB, which is always built-in on arm64.

SDHCI controllers can use 32-bit DMA or 64-bit DMA, however even with
64-bit DMA it is possible to restrict the DMA addresses to 32-bits
by setting a 32-bit DMA mask.

If the host controller capabilities indicate support for 64-bit DMA
but you want the driver to use 32-bit DMA, set SDHCI_QUIRK2_BROKEN_64_BIT_DMA.

However, if you want to use 64-bit DMA with only 32-bit DMA addresses 
you can instead implement sdhci host op ->set_dma_mask() and in that
function dma_set_mask_and_coherent(dev, DMA_BIT_MASK(32))

I would not expect an standard SDHCI to be broken like this any more,
and if it is, I think the SDHCI_QUIRK2_BROKEN_64_BIT_DMA quirk is
more appropriate than overriding the dma_set_mask() operation.

What sometimes happens though is that the SoC integration itself is
broken, and a 64-bit capable DMA master like SDHCI is connected
to a 32-bit bus. In this case the DMA limitation should be
described in the device tree, using the "dma-ranges" property
of the broken bus node. The SDHCI code then still sets the correct
64-bit DMA mask according for the device, but the dma_map_single()
still uses an swiotlb bounce buffer or the IOMMU to work around
the bus restriction.

quoted

 To resolve this hardware limitation, we implement a bounce buffer allocated via >> dma_alloc_coherent() to satisfy DMA addressing constraints.

The bounce buffer should not be needed to satisfy DMA addressing
constraints.  It is used when SDHCI ADMA (scatter/gather) is broken.

I wonder if the actual problem here is not the addressing limit
but instead the coherency protocol. If the DMA master is cache
coherent but listed as non-coherent in DT, or vice versa, there will
be data corruption for all addresses, and using a dma_alloc_coherent()
bounce buffer may hide this.

       Arnd