Re: [PATCH v2 06/18] x86/kasan/64: Teach KASAN about the cpu_entry_area

[Andy Lutomirski]

On Thu, Nov 23, 2017 at 2:08 AM, Andrey Ryabinin
<aryabinin@xxxxxxxxxxxxx> wrote:
>
>
> On 11/22/2017 06:22 PM, Andy Lutomirski wrote:
>> On Wed, Nov 22, 2017 at 1:05 AM, Andrey Ryabinin
>> <aryabinin@xxxxxxxxxxxxx> wrote:
>>>
>>>
>>> On 11/22/2017 07:44 AM, Andy Lutomirski wrote:
>>>> The cpu_entry_area will contain stacks.  Make sure that KASAN has
>>>> appropriate shadow mappings for them.
>>>>
>>>> Cc: Andrey Ryabinin <aryabinin@xxxxxxxxxxxxx>
>>>> Cc: Alexander Potapenko <glider@xxxxxxxxxx>
>>>> Cc: Dmitry Vyukov <dvyukov@xxxxxxxxxx>
>>>> Cc: kasan-dev@xxxxxxxxxxxxxxxx
>>>> Signed-off-by: Andy Lutomirski <luto@xxxxxxxxxx>
>>>> ---
>>>>  arch/x86/mm/kasan_init_64.c | 9 ++++++++-
>>>>  1 file changed, 8 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/arch/x86/mm/kasan_init_64.c b/arch/x86/mm/kasan_init_64.c
>>>> index 99dfed6dfef8..43d376687315 100644
>>>> --- a/arch/x86/mm/kasan_init_64.c
>>>> +++ b/arch/x86/mm/kasan_init_64.c
>>>> @@ -330,7 +330,14 @@ void __init kasan_init(void)
>>>>                             early_pfn_to_nid(__pa(_stext)));
>>>>
>>>>       kasan_populate_zero_shadow(kasan_mem_to_shadow((void *)MODULES_END),
>>>> -                     (void *)KASAN_SHADOW_END);
>>>> +                                kasan_mem_to_shadow((void *)(__fix_to_virt(FIX_CPU_ENTRY_AREA_BOTTOM))));
>>>> +
>>>> +     kasan_populate_shadow((unsigned long)kasan_mem_to_shadow((void *)(__fix_to_virt(FIX_CPU_ENTRY_AREA_BOTTOM))),
>>>> +                           (unsigned long)kasan_mem_to_shadow((void *)(__fix_to_virt(FIX_CPU_ENTRY_AREA_TOP) + PAGE_SIZE)),
>>>
>>> What's '+ PAGE_SIZE' for?
>>>
>>
>> __fix_to_virt(..._TOP) returns the address of the *bottom* of the last
>> cpu_entry_area page.  +PAGE_SIZE returns one past the end of the
>> region, which I assume is the correct thing to pass.
>>
>
> Right.
>
> Perhaps, it would be better to use variables, just avoid such awfully long lines, I mean like this:
>         fixmap_shadow_start = (void *)__fix_to_virt(FIX_CPU_ENTRY_AREA_BOTTOM);
>         fixmap_shadow_start = kasan_mem_to_shadow(fixmap_shadow_start);
>
>         fixmap_shadow_end = (void *)__fix_to_virt(FIX_CPU_ENTRY_AREA_TOP) + PAGE_SIZE;
>         fixmap_shadow_end = kasan_mem_to_shadow(fixmap_shadow_end);
>
>         kasan_populate_zero_shadow(kasan_mem_to_shadow((void *)MODULES_END),
>                                    fixmap_shadow_start);
>
>         kasan_populate_shadow((unsigned long)fixmap_shadow_start,
>                         (unsigned long)fixmap_shadow_end,
>                         0);
>
> I'm also thinking that we should change kasan_populate_shadow() to take void* instead of 'unsigned long'
> to avoid those casts.

I did something similar, but I left the kasan_mem_to_shadow in for
consistency with the rest.

I think the real way to clean this up is like this:

struct kasan_range {
  void *start;
  void *end;
};

struct kasan_range ranges[] = {
 { range 1 },
 { range 2 },
};

sort(range, ARRAY_SIZE(ranges), sizeof(ranges[0]), ...);

last_end = PAGE_OFFSET;  /* or whatever is right */

for (i = 0; i < ARRAY_SIZE(ranges); i++) [
  WARN_ON(ranges[i].start < last_end);

  if (ranges[i].start > last_end)
    kasan_populate_zero_shadow(...);
  kasan_populate_shadow(...);
 last_end = ranges[i].end;
}

kasan_populate_zero_shadow(last_end, the real end);

Then the code doesn't need to duplicate each boundary or to hardcode
the order in which things appear.