As it turns out, arm64 deviates from other architectures in the way it
maps the VMALLOC region: on most (all?) other architectures, it resides
strictly above the kernel's direct mapping of DRAM, but on arm64, this
is the other way around. For instance, for a 48-bit VA configuration,
we have

  modules : 0xffff000000000000 - 0xffff000008000000   (   128 MB)
  vmalloc : 0xffff000008000000 - 0xffff7dffbfff0000   (129022 GB)
  ...
  vmemmap : 0xffff7e0000000000 - 0xffff800000000000   (  2048 GB maximum)
            0xffff7e0000000000 - 0xffff7e0003ff0000   (    63 MB actual)
  memory  : 0xffff800000000000 - 0xffff8000ffc00000   (  4092 MB)

This has mostly gone unnoticed until now, but it does appear that it
breaks an assumption in the kmem read/write code, which does something
like

  if (p < (unsigned long) high_memory) {
    ... use straight copy_[to|from]_user() using p as virtual address ...
  }
  ...
  if (count > 0) {
    ... use vread/vwrite for accesses past high_memory ...
  }

The first condition will inadvertently hold for the VMALLOC region if
VMALLOC_START < PAGE_OFFSET [which is the case on arm64], but the read
or write will subsequently fail the virt_addr_valid() check, resulting
in a -ENXIO return value.

Given how kmem seems to be living in borrowed time anyway, and given
the fact that nobody noticed that the read/write interface is broken
on arm64 in the first place, let's not bother trying to fix it, but
simply disable the /dev/kmem interface entirely for arm64.
Acked-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Will Deacon <will.deacon@arm.com>