Re: [PATCH 1/2] mm/zsmalloc.c: check encoded object value overflow for PAE

[Rafael David Tinoco <rafael.tinoco@linaro.org>]

On Thu, Oct 25, 2018 at 2:29 AM, Sergey Senozhatsky
<sergey.senozhatsky.work@gmail.com> wrote:
> On (10/24/18 22:27), Rafael David Tinoco wrote:
>>  static unsigned long location_to_obj(struct page *page, unsigned int obj_idx)
>>  {
>> -     unsigned long obj;
>> +     unsigned long obj, pfn;
>> +
>> +     pfn = page_to_pfn(page);
>> +
>> +     if (unlikely(OBJ_OVERFLOW(pfn)))
>> +             BUG();
>
> The trend these days is to have less BUG/BUG_ON-s in the kernel.
>
>         -ss

For this case, IMHO, it is worth.

It will avoid a investigation like:
https://bugs.linaro.org/show_bug.cgi?id=3765#c7 and and #c8, where I
had to poison slab allocation - to force both zs_handle and zspage
slabs not to be merged - and to make sure the zspage slab had a good
magic number AND to identify why the bad paging request happened.

If this happens again, for any other arch supporting PAE that does not
declare MAX_POSSIBLE_PHYSMEM_BITS or MAX_PHYSMEM_BITS appropriately,
the kernel will panic, no matter what, by the time it reaches
obj_to_location(). Things can be more complicated about declarations
for PAE if we consider ARM can declare MAX_PHYSMEM_BITS differently in
arch/arm/mach-XXX and/or, for this case, when having, or not SPARSEMEM
set (if I had SPARSEMEM set I would not face this, for example).

If this occurs, the kernel will panic, no matter what, by the time it
reaches obj_to_location()... so why not to BUG() here and let user to
know exactly where it panic-ed and why ? Other option would be to
WARN() here and let it panic naturally because of bad paging request
in a very near future... please advise.

Thanks,
Best Rgds
-Rafael