Re: [PATCH] kasan: fix the missing underflow in memmove and memcpy with CONFIG_KASAN_GENERIC=y

[Walter Wu]

On Mon, 2019-09-30 at 10:57 +0200, Marc Gonzalez wrote:
> On 30/09/2019 06:36, Walter Wu wrote:
> 
> >  bool check_memory_region(unsigned long addr, size_t size, bool write,
> >                                 unsigned long ret_ip)
> >  {
> > +       if (long(size) < 0) {
> > +               kasan_report_invalid_size(src, dest, len, _RET_IP_);
> > +               return false;
> > +       }
> > +
> >         return check_memory_region_inline(addr, size, write, ret_ip);
> >  }
> 
> Is it expected that memcpy/memmove may sometimes (incorrectly) be passed
> a negative value? (It would indeed turn up as a "large" size_t)
> 
> IMO, casting to long is suspicious.
> 
> There seem to be some two implicit assumptions.
> 
> 1) size >= ULONG_MAX/2 is invalid input
> 2) casting a size >= ULONG_MAX/2 to long yields a negative value
> 
> 1) seems reasonable because we can't copy more than half of memory to
> the other half of memory. I suppose the constraint could be even tighter,
> but it's not clear where to draw the line, especially when considering
> 32b vs 64b arches.
> 
> 2) is implementation-defined, and gcc works "as expected" (clang too
> probably) https://gcc.gnu.org/onlinedocs/gcc/Integers-implementation.html
> 
> A comment might be warranted to explain the rationale.
> Regards.

Thanks for your suggestion.
Yes, It is passed a negative value issue in memcpy/memmove/memset.
Our current idea should be assumption 1 and only consider 64b arch,
because KASAN only supports 64b. In fact, we really can't use so much
memory in 64b arch. so assumption 1 make sense.