Nicholas Piggin's on April 3, 2020 9:05 pm: > Christophe Leroy's on April 3, 2020 8:31 pm: >> >> >> Le 03/04/2020 à 11:35, Nicholas Piggin a écrit : >>> There is no need to allow user accesses when probing kernel addresses. >> >> I just discovered the following commit >> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=75a1a607bb7e6d918be3aca11ec2214a275392f4 >> >> This commit adds probe_kernel_read_strict() and probe_kernel_write_strict(). >> >> When reading the commit log, I understand that probe_kernel_read() may >> be used to access some user memory. Which will not work anymore with >> your patch. > > Hmm, I looked at _strict but obviously not hard enough. Good catch. > > I don't think probe_kernel_read() should ever access user memory, > the comment certainly says it doesn't, but that patch sort of implies > that they do. > > I think it's wrong. The non-_strict maybe could return userspace data to > you if you did pass a user address? I don't see why that shouldn't just > be disallowed always though. > > And if the _strict version is required to be safe, then it seems like a > bug or security issue to just allow everyone that doesn't explicitly > override it to use the default implementation. > > Also, the way the weak linkage is done in that patch, means parisc and > um archs that were previously overriding probe_kernel_read() now get > the default probe_kernel_read_strict(), which would be wrong for them. The changelog in commit 75a1a607bb7 makes it a bit clearer. If the non-_strict variant is used on non-kernel addresses, then it might not return -EFAULT or it might cause a kernel warning. The _strict variant is supposed to be usable with any address and it will return -EFAULT if it was not a valid and mapped kernel address. The non-_strict variant can not portably access user memory because it uses KERNEL_DS, and its documentation says its only for kernel pointers. So powerpc should be fine to run that under KUAP AFAIKS. I don't know why the _strict behaviour is not just made default, but the implementation of it does seem to be broken on the archs that override the non-_strict variant. Thanks, Nick
