On Wed, 9 Oct 2024 at 22:19, Andrey Konovalov <andreyknvl@xxxxxxxxx> wrote: [...] > Please add a comment here explaining why we only check > copy_to_kernel_nofault and not copy_from_kernel_nofault (is this > because we cannot add KASAN instrumentation to > copy_from_kernel_nofault?). Just to clarify: Unless we can prove that there won't be any false positives, I proposed to err on the side of being conservative here. The new way of doing it after we already checked that the accessed location is on a faulted-in page may be amenable to also KASAN instrumentation, but you can also come up with cases that would be a false positive: e.g. some copy_from_kernel_nofault() for a large range, knowing that if it accesses bad memory at least one page is not faulted in, but some initial pages may be faulted in; in that case there'd be some error handling that then deals with the failure. Again, this might be something that an eBPF program could legally do. On the other hand, we may want to know if we are leaking random uninitialized kernel memory with KMSAN to avoid infoleaks. Only copy_to_kernel_nofault should really have valid memory, otherwise we risk corrupting the kernel. But these checks should only happen after we know we're accessing faulted-in memory, again to avoid false positives.
