On Mon, Nov 25, 2024 at 10:50:48AM +0000, Mark Rutland wrote: > On Mon, Nov 25, 2024 at 10:54:49AM +0100, Lukas Wunner wrote: > > Other arches do not seem to be concerned about this and > > let virt_addr_valid() return true for the kernel image. > > It's not clear why arm64 is special and needs to return false. > > > > However, I agree there's hardly ever a reason to DMA from/to the > > .text section. From a security perspective, constraining this to > > .rodata seems reasonable to me and I'll be happy to amend the patch > > to that effect if that's the consensus. > > Instead, can we update the test to use lm_alias() on the symbols in > question? That'll convert a kernel image address to its linear map > alias, and then that'll work with virt_addr_valid(), virt_to_phys(), > etc. Do you mean that sg_set_buf() should pass the address to lm_alias() if it points into the kernel image? That would require a helper to determine whether it's a kernel image address or not. It seems we do not have such a cross-architecture helper (but maybe I'm missing something). (I am adding an arm64-specific one in the proposed patch.) So this doesn't look like a viable approach. Also, I'd expect pushback against an sg_set_buf() change which is only necessary to accommodate arm64. I'd expect the obvious question to be asked, which is why arm64's virt_addr_valid() can't behave like any other architecture's. And honestly I wouldn't know what to answer. Thanks, Lukas
