Hi, I am suffering from performance issue similar to that is supposed to be fixed by 3e0ac6717b0a ("[media] media: Fix a UVC performance problem on systems with non-coherent DMA") https://github.com/torvalds/linux/commit/3e0ac6717b0a2593243ef3850025a42cb19... The patch operates with CONFIG_DMA_NONCOHERENT compile-time option which is present only for mips and sh architectures. At the same time, some ARM platforms are also DMA non-coherent that leads to the similar memcpy() performance issues as mentioned in the patch. However, we learn whether we run with coherent or non-coherent DMA at run-time while parsing FDT in case of ARM platforms. This happens in of_dma_configure() function. I would like to find a way to re-implement the patch for the case of ARM platforms, but unfortunately I could not find preferred way to check DMA coherency in run-time. Could you please advice? p.s. Frankly speaking, the patch looks to be redundant for me. I have no idea why DMA coherency could be really required for USB Isochronous input buffer. At the host side there is only reading operation, that means that single barrier would have been sufficient here.