On Wednesday, June 29, 2016 9:33:54 AM CEST Timur Tabi wrote: > Arnd Bergmann wrote: > > If the ranges property lists the bus as dma capable for only the > > lower 32 bits, then dma_set_mask_and_coherent(dev, DMA_BIT_MASK(64)); > > should fail, otherwise dma_alloc_coherent() will return an invalid > > memory area. > > That seems wrong. dma_alloc_coherent() should be smart enough to > restrict itself to the the dma-ranges property. Isn't that why the > property exists? When dma_alloc_coherent() looks for memory, it should > knows it has to create a 32-bit address. That's why we have ZONE_DMA. No, dma_alloc_coherent() is documented to use the dma_mask as its reference, it's supposed to be independent of the underlying bus. dma-ranges is just how we communicate the limitation of the bus to the kernel, but relying on dma-ranges itself would fail to consider drivers that impose additional limitations, i.e. when a device needs a smaller mask and calls 'dma_set_mask(dev, DMA_BIT_MASK(24))' or something like that. > > Another twist is how arm64 currently uses SWIOTLB unconditionally: > > As long as SWIOTLB (or iommu) is enabled, dma_set_mask_and_coherent() > > should succeed for any mask(), but not actually update the mask of the > > device to more than the bus can handle. > > That just seems like a bug in ARM64 SWIOTLB. SWIOTLB should inject > itself when the driver tries to map memory outside of its DMA range. Again, the dma mask is how swiotlb_map_*() finds whether a page needs a bounce buffer or not, so it has to be set to whatever the device can address. > Without SWIOTLB/IOMMU, dma_alloc_coherent() should be aware of the > platform-specific limitations of each device and ensure that it only > allocates memory that conforms *all* limitations. For example, if the > platform is capable of 64-bit DMA, but a legacy device can only handle > 32-bit bus addresses, then the driver should do this: > > dma_set_mask_and_coherent(dev, DMA_BIT_MASK(32)) That's also not how it works: each device starts out with a 32-bit mask, because that's what historically all PCI devices can do. If a device is 64-bit DMA capable, it can extend the mask by passing DMA_BIT_MASK(64) (or whatever it can support), and the platform code checks if that's possible. Arnd -- To unsubscribe from this list: send the line "unsubscribe devicetree" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
