On Fri, May 13, 2022 at 12:37:24PM +0200, Christoph Hellwig wrote: > > #regzbot introduced: f5ff79fddf0efecca538046b5cc20fb3ded2 > > Well, this just uncovered an existing bug in the driver. You can not > just dma map memory returned from dma_alloc_coherent, and this driver > would already get vmalloc memory on arm/arm64 platforms anyway, we > now just do the same on x86 as well. >From a quick look through the trace it seems somehow usb_hcd_map_urb_for_dma tries to create another DMA mapping for a buffer allocated from usb_alloc_coherent that uses a dma coherent allocation below. It really needs to use the dma address returned from usb_alloc_coherent instead of trying to map the URB again. But I don't have the slightest idea of why that only happens for this particular setup, and I really need some helpe from the usb folks to untangle it
