On Fri, 2019-05-17 at 13:29:58 UTC, Michael Ellerman wrote:
> From: "Aneesh Kumar K.V" <aneesh.ku...@linux.ibm.com>
>
> Accesses by userspace to random addresses outside the user or kernel
> address range will generate an SLB fault. When we handle that fault we
> classify the effective address into several classes, eg. user, kernel
> linear, kernel virtual etc.
>
> For addresses that are completely outside of any valid range, we
> should not insert an SLB entry at all, and instead immediately an
> exception.
>
> In the past this was handled in two ways. Firstly we would check the
> top nibble of the address (using REGION_ID(ea)) and that would tell us
> if the address was user (0), kernel linear (c), kernel virtual (d), or
> vmemmap (f). If the address didn't match any of these it was invalid.
>
> Then for each type of address we would do a secondary check. For the
> user region we check against H_PGTABLE_RANGE, for kernel linear we
> would mask the top nibble of the address and then check the address
> against MAX_PHYSMEM_BITS.
>
> As part of commit 0034d395f89d ("powerpc/mm/hash64: Map all the kernel
> regions in the same 0xc range") we replaced REGION_ID() with
> get_region_id() and changed the masking of the top nibble to only mask
> the top two bits, which introduced a bug.
>
> Addresses less than (4 << 60) are still handled correctly, they are
> either less than (1 << 60) in which case they are subject to the
> H_PGTABLE_RANGE check, or they are correctly checked against
> MAX_PHYSMEM_BITS.
>
> However addresses from (4 << 60) to ((0xc << 60) - 1), are incorrectly
> treated as kernel linear addresses in get_region_id(). Then the top
> two bits are cleared by EA_MASK in slb_allocate_kernel() and the
> address is checked against MAX_PHYSMEM_BITS, which it passes due to
> the masking. The end result is we incorrectly insert SLB entries for
> those addresses.
>
> That is not actually catastrophic, having inserted the SLB entry we
> will then go on to take a page fault for the address and at that point
> we detect the problem and report it as a bad fault.
>
> Still we should not be inserting those entries, or treating them as
> kernel linear addresses in the first place. So fix get_region_id() to
> detect addresses in that range and return an invalid region id, which
> we cause use to not insert an SLB entry and directly report an
> exception.
>
> Fixes: 0034d395f89d ("powerpc/mm/hash64: Map all the kernel regions in the
> same 0xc range")
> Signed-off-by: Aneesh Kumar K.V <aneesh.ku...@linux.ibm.com>
> [mpe: Drop change to EA_MASK for now, rewrite change log]
> Signed-off-by: Michael Ellerman <m...@ellerman.id.au>
Applied to powerpc fixes.
https://git.kernel.org/powerpc/c/c179976cf4cbd2e65f29741d5bc07ccf
cheers
