Re: [PATCH RESEND 1/4] uaccess: Add user_read_access_begin/end and user_write_access_begin/end

[Catalin Marinas <catalin.marinas@arm.com>]

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