mmc: tegra: Implement ->set_dma_mask()
The SDHCI controller on Tegra186 supports 40-bit addressing, which is usually enough to address all of system memory. However, if the SDHCI controller is behind an IOMMU, the address space can go beyond. This happens on Tegra186 and later where the ARM SMMU has an input address space of 48 bits. If the DMA API is backed by this ARM SMMU, the top- down IOVA allocator will cause IOV addresses to be returned that the SDHCI controller cannot access. Unfortunately, prior to the introduction of the ->set_dma_mask() host operation, the SDHCI core would set either a 64-bit DMA mask if the controller claimed to support 64-bit addressing, or a 32-bit DMA mask otherwise. Since the full 64 bits cannot be addressed on Tegra, this had to be worked around in commit 68481a7e ("mmc: tegra: Mark 64 bit dma broken on Tegra186") by setting the SDHCI_QUIRK2_BROKEN_64_BIT_DMA quirk, which effectively restricts the DMA mask to 32 bits. One disadvantage of this is that dma_map_*() APIs will now try to use the swiotlb to bounce DMA to addresses beyond of the controller's DMA mask. This in turn caused degraded performance and can lead to situations where the swiotlb buffer is exhausted, which in turn leads to DMA transfers to fail. With the recent introduction of the ->set_dma_mask() host operation, this can now be properly fixed. For each generation of Tegra, the exact supported DMA mask can be configured. This kills two birds with one stone: it avoids the use of bounce buffers because system memory never exceeds the addressable memory range of the SDHCI controllers on these devices, and at the same time when an IOMMU is involved, it prevents IOV addresses from being allocated beyond the addressible range of the controllers. Since the DMA mask is now properly handled, the 64-bit DMA quirk can be removed. Signed-off-by:Nicolin Chen <nicoleotsuka@gmail.com> [treding@nvidia.com: provide more background in commit message] Tested-by:
Please register or sign in to comment