On Wed, 20 Sep 2017, Leonard Crestez wrote: > On Mon, 2017-09-18 at 11:08 -0700, Stefano Stabellini wrote: > On Fri, 15 Sep 2017, Greg KH wrote: > > On Thu, Sep 14, 2017 at 04:23:05PM -0700, Stefano Stabellini wrote: > > > Hi all, > > > > > > We are getting reports from Xen on ARM users about DMA issues. The > > > problem is that the commit below > > > (7e91c7df29b5e196de3dc6f086c8937973bd0b88) is necessary to support mmap > > > on Xen on ARM. It is self-contained and doesn't affect anything outside > > > of Xen on ARM, so I think is a good candidate for backporting. It went > > > upstream in 4.11. > > > > But it's a new feature, right? How does that fit the stable kernel > > rules? > > > It implements a previously unimplemented function (mmap), although it > calls the generic functions to do it. Yes, I agree with you that it > can be classified as a new feature. If that is against the stable kernel > rules, then please discard this request. > > > FYI the reason why it didn't raise a flag in my mind is that users > reported something like "unhandled alignment fault (11) at > 0xffffa6048080, esr 0x92000061", which really looks more like a bug. > > I am the one who reported this, on the #xenarm IRC channel. Thank you for jumping into this thread. > Not implementing mmap in dma_map_ops means that dma_common_mmap is > called by dma_map_attrs as a fallback. The end result is not something > like -ENOSYS but what seem to be corrupt mappings. > > However I agree that backporting might be excessive. I ran into this by > experimenting with using a GPU from dom0. It seems reasonable to get > kernel crashes if you try this kind of stuff. > > This patch results in calling __swiotlb_mmap instead of > dma_common_mmap. I don't know the implementation details of the DMA api > but the interesting difference between these paths seems to be the way > pfn is fetched (from dma_addr instead of the kernel virt addr). Yes, on ARM and ARM64 dma_map_ops functions can return pages for which virt_to_page doesn't work as expected (for example on ARM alloc_coherent returns an ioremap'ped virtual address, I don't remember the details of the ARM64 implementation right now). This is why the dma_map_ops functions are implemented by looking up the physical address from the dma address.