On Thu, Jan 25, 2024 at 10:47:26AM +0100, Arnd Bergmann wrote: > On Thu, Jan 25, 2024, at 09:50, Linus Walleij wrote: > > The MMC core sets BLK_BOUNCE_HIGH for devices where dma_mask > > is unassigned. > > > > For the majority of MMC hosts this path is never taken: the > > OF core will unconditionally assign a 32-bit mask to any > > OF device, and most MMC hosts are probed from device tree, > > see drivers/of/platform.c: > > > > of_platform_device_create_pdata() > > dev->dev.coherent_dma_mask = DMA_BIT_MASK(32); > > if (!dev->dev.dma_mask) > > dev->dev.dma_mask = &dev->dev.coherent_dma_mask; > > > > of_amba_device_create() > > dev->dev.coherent_dma_mask = DMA_BIT_MASK(32); > > dev->dev.dma_mask = &dev->dev.coherent_dma_mask; > > > > MMC devices that are probed from ACPI or PCI will likewise > > have a proper dma_mask assigned. > > > > The only remaining devices that could have a blank dma_mask > > are platform devices instantiated from board files. > > > > These are mostly used on systems without CONFIG_HIGHMEM > > enabled which means the block layer will not bounce, and in > > the few cases where it is enabled it is not used anyway: > > for example some OMAP2 systems such as Nokia n800/n810 will > > create a platform_device and not assign a dma_mask, however > > they do not have any highmem, so no bouncing will happen > > anyway: the block core checks if max_low_pfn >= max_pfn > > and this will always be false. > > > > Should it turn out there is a platform_device with blank > > DMA mask actually using CONFIG_HIGHMEM somewhere out there > > we should set dma_mask for it, not do this trickery. > > > > Signed-off-by: Linus Walleij <linus.walleij@xxxxxxxxxx> > > Acked-by: Arnd Bergmann <arnd@xxxxxxxx> > > I think it's worth mentioning the cb710 example here, which > uses a platform device as a child of a PCI device and > does not assign a DMA mask nor use DMA. > > This one will see a change in behavior, meaning that the > blockdev buffers are no longer bounced. As far as I can > tell, this is fine because the driver appears to correctly > use the sg_iter infrastructure for mapping data pages, > but it would be good to have this confirmed by > Michał Mirosław because this code path has probably never > been tested without BLK_BOUNCE_HIGH. Hi, this driver doesn't do DMA at all, so having DMA mask set or not it should be good as long as the CPU can read/write the buffers. Best Regards Michał Mirosław
