powerpc/mm/radix: Make Radix require HUGETLB_PAGE

Joel reported weird crashes using skiroot_defconfig, in his case we
jumped into an NX page:

  kernel tried to execute exec-protected page (c000000002bff4f0) - exploit attempt? (uid: 0)
  BUG: Unable to handle kernel instruction fetch
  Faulting instruction address: 0xc000000002bff4f0

Looking at the disassembly, we had simply branched to that address:

  c000000000c001bc  49fff335    bl     c000000002bff4f0

But that didn't match the original kernel image:

  c000000000c001bc  4bfff335    bl     c000000000bff4f0 <kobject_get+0x8>

When STRICT_KERNEL_RWX is enabled, and we're using the radix MMU, we
call radix__change_memory_range() late in boot to change page
protections. We do that both to mark rodata read only and also to mark
init text no-execute. That involves walking the kernel page tables,
and clearing _PAGE_WRITE or _PAGE_EXEC respectively.

With radix we may use hugepages for the linear mapping, so the code in
radix__change_memory_range() uses eg. pmd_huge() to test if it has
found a huge mapping, and if so it stops the page table walk and
changes the PMD permissions.

However if the kernel is built without HUGETLBFS support, pmd_huge()
is just a #define that always returns 0. That causes the code in
radix__change_memory_range() to incorrectly interpret the PMD value as
a pointer to a PTE page rather than as a PTE at the PMD level.

Unfortunately the combination of _PAGE_PTE and _PAGE_PRESENT in the
high bits of the PMD entry give us 0xc in the top nibble which means
the PMD entry happens to look like a valid pointer into the linear
mapping.

We can see this using `dv` in xmon:

  0:mon> dv c000000000000000
  pgd  @ 0xc000000001740000
  pgdp @ 0xc000000001740000 = 0x80000000ffffb009
  pudp @ 0xc0000000ffffb000 = 0x80000000ffffa009
  pmdp @ 0xc0000000ffffa000 = 0xc00000000000018f	<- not a pointer
  ptep @ 0xc000000000000100 = 0xa64bb17da64ab07d	<- kernel text

The end result is we treat the value at 0xc000000000000100 as a PTE
and clear _PAGE_WRITE or _PAGE_EXEC, potentially corrupting the code
at that address.

In Joel's specific case we cleared the sign bit in the offset of the
branch, causing a backward branch to turn into a forward branch which
caused us to branch into a non-executable page. However the exact
nature of the crash depends on kernel version, compiler version, and
other factors.

We need to fix radix__change_memory_range() to not use accessors that
depend on HUGETLBFS, but we also have radix memory hotplug code that
uses pmd_huge() etc that will also need fixing. So for now just
disallow the broken combination of Radix with HUGETLBFS disabled.

The only defconfig we have that is affected is skiroot_defconfig, so
turn on HUGETLBFS there so that it still gets Radix.

Fixes: 566ca99af026 ("powerpc/mm/radix: Add dummy radix_enabled()")
Cc: stable@vger.kernel.org # v4.7+
Reported-by: Joel Stanley <joel@jms.id.au>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
---
 arch/powerpc/configs/skiroot_defconfig | 1 +
 arch/powerpc/platforms/Kconfig.cputype | 2 +-
 2 files changed, 2 insertions(+), 1 deletion(-)

Michael Ellerman <mpe@ellerman.id.au> writes:
> Joel reported weird crashes using skiroot_defconfig, in his case we
> jumped into an NX page:
>
>   kernel tried to execute exec-protected page (c000000002bff4f0) - exploit attempt? (uid: 0)
>   BUG: Unable to handle kernel instruction fetch
>   Faulting instruction address: 0xc000000002bff4f0
>
> Looking at the disassembly, we had simply branched to that address:
>
>   c000000000c001bc  49fff335    bl     c000000002bff4f0
>
> But that didn't match the original kernel image:
>
>   c000000000c001bc  4bfff335    bl     c000000000bff4f0 <kobject_get+0x8>
>
> When STRICT_KERNEL_RWX is enabled, and we're using the radix MMU, we
> call radix__change_memory_range() late in boot to change page
> protections. We do that both to mark rodata read only and also to mark
> init text no-execute. That involves walking the kernel page tables,
> and clearing _PAGE_WRITE or _PAGE_EXEC respectively.
>
> With radix we may use hugepages for the linear mapping, so the code in
> radix__change_memory_range() uses eg. pmd_huge() to test if it has
> found a huge mapping, and if so it stops the page table walk and
> changes the PMD permissions.
>
> However if the kernel is built without HUGETLBFS support, pmd_huge()
> is just a #define that always returns 0. That causes the code in
> radix__change_memory_range() to incorrectly interpret the PMD value as
> a pointer to a PTE page rather than as a PTE at the PMD level.
>
> Unfortunately the combination of _PAGE_PTE and _PAGE_PRESENT in the
> high bits of the PMD entry give us 0xc in the top nibble which means
> the PMD entry happens to look like a valid pointer into the linear
> mapping.

Aneesh points out this paragraph is confusing. Those bits only make the
pointer *look* valid, it doesn't actually make any difference to the
code, as we mask and re-add those bits anyway.

I'll drop this paragraph when committing.

cheers