On Fri, Apr 03, 2020 at 01:58:31AM +0100, Al Viro wrote: > On Thu, Apr 02, 2020 at 11:35:57AM -0700, Kees Cook wrote: > > Yup, I think it's a weakness of the ARM implementation and I'd like to > > not extend it further. AFAIK we should never nest, but I would not be > > surprised at all if we did. > > > > If we were looking at a design goal for all architectures, I'd like > > to be doing what the public PaX patchset did for their memory access > > switching, which is to alarm if calling into "enable" found the access > > already enabled, etc. Such a condition would show an unexpected nesting > > (like we've seen with similar constructs with set_fs() not getting reset > > during an exception handler, etc etc). > > FWIW, maybe I'm misreading the ARM uaccess logics, but... it smells like > KERNEL_DS is somewhat more dangerous there than on e.g. x86. > > Look: with CONFIG_CPU_DOMAINS, set_fs(KERNEL_DS) tells MMU to ignore > per-page permission bits in DOMAIN_KERNEL (i.e. for kernel address > ranges), allowing them even if they would normally be denied. We need > that for actual uaccess loads/stores, since those use insns that pretend > to be done in user mode and we want them to access the kernel pages. > But that affects the normal loads/stores as well; unless I'm misreading > that code, it will ignore (supervisor) r/o on a page. And that's not > just for the code inside the uaccess blocks; *everything* done under > KERNEL_DS is subject to that. That's correct. Luckily this only affects ARMv5 and earlier. From ARMv6 onwards, CONFIG_CPU_USE_DOMAINS is no longer selected and the uaccess instructions are just plain ldr/str. Russell should know the details on whether there was much choice. Since the kernel was living in the linear map with full rwx permissions, the KERNEL_DS overriding was probably not a concern and the ldrt/strt for uaccess deemed more secure. We also have weird permission setting pre-ARMv6 (or rather v6k) where RO user pages are writable from the kernel with standard str instructions (breaking CoW). I don't recall whether it was a choice made by the kernel or something the architecture enforced. The vectors page has to be kernel writable (and user RO) to store the TLS value in the absence of a TLS register but maybe we could do this via the linear alias together with the appropriate cache maintenance. >From ARMv6, the domain overriding had the side-effect of ignoring the XN bit and causing random instruction fetches from ioremap() areas. So we had to remove the domain switching. We also gained a dedicated TLS register. > Why do we do that (modify_domain(), that is) inside set_fs() and not > in uaccess_enable() et.al.? I think uaccess_enable() could indeed switch the kernel domain if KERNEL_DS is set and move this out of set_fs(). It would reduce the window the kernel domain permissions are overridden. Anyway, uaccess_enable() appeared much later on arm when Russell introduced PAN (SMAP) like support by switching the user domain. -- Catalin
