arm64: mm: bug around swiotlb_dma_ops

From: Nikita Yushchenko
Date: Thu Dec 15 2016 - 11:20:19 EST


Per Documentation/DMA-API-HOWTO.txt, driver of device capable of 64-bit
DMA addressing, should call dma_set_mask_and_coherent(dev,
DMA_BIT_MASK(64)) and if that succeeds, assume that 64-bit DMA
addressing is available.

This behaves incorrectly on arm64 system (Renesas r8a7795-h3ulcb) here.

- Device (NVME SSD) has it's dev->archdata.dma_ops set to swiotlb_dma_ops.

- swiotlb_dma_ops.dma_supported is set to swiotlb_dma_supported():

int swiotlb_dma_supported(struct device *hwdev, u64 mask)
return phys_to_dma(hwdev, io_tlb_end - 1) <= mask;

this definitely returns true for mask=DMA_BIT_MASK(64) since that is
maximum possible 64-bit value.

- Thus device dma_mask is unconditionally updated, and
dma_set_mask_and_coherent() succeeds.

- Later, __swiotlb_map_page() / __swiotlb_map_sg_attr() will consult
this updated mask, and return high addresses as valid DMA addresses.

Thus recommended dma_set_mask_and_coherent() call, instead of checking
if platform supports 64-bit DMA addressing, unconditionally enables
64-bit DMA addressing. In case of device actually can't do DMA to 64-bit
addresses (e.g. because of limitations in PCIe controller), this breaks
things. This is exactly what happens here.

Not sure what is proper fix for this though.