Re: [PATCH] find: Do not read beyond variable boundaries on small sizes

[Yury Norov <yury.norov@gmail.com>]

On Fri, Dec 03, 2021 at 08:37:59AM -0800, Kees Cook wrote:
> 
> 
> On December 3, 2021 4:30:35 AM PST, Andy Shevchenko <andriy.shevchenko@linux.intel.com> wrote:
> >On Fri, Dec 03, 2021 at 02:08:46AM -0800, Kees Cook wrote:
> >> It's common practice to cast small variable arguments to the find_*_bit()
> >
> >It's a bad practice and should be fixed accordingly, no?
> 
> There's an argument to be made that the first arg should be void * but that's a pretty invasive change at this point (and orthogonal to this fix).

What for? To save at most 7 bytes of alignment overhead for bitmaps
like char bitmap[sizeof(unsigned long) + 1]?
 
> I'd be happy to send a treewide change for that too, if folks wanted?

For small arrays of bits that are fraction of machine word we have
ffs/fls/ffz. For long bitmaps, the alignment overhead is not that
important - at least nobody complained.

If we convert bitmaps to void*, it would mean that we'd handle tails
just like you did in this patch. The __find_bits_deref()-style function
should be also called from each lib/bitmap.c function together with
store() analogue, and overall impact would barely be positive.

Char-aligned bitmaps would be traversed less efficient than word-aligned
on most architectures, and we'll have the same problems that memcpy() has. 

Thanks,
Yury

> >> helpers to unsigned long and then use a size argument smaller than
> >> sizeof(unsigned long):
> >> 
> >> 	unsigned int bits;
> >> 	...
> >> 	out = find_first_bit((unsigned long *)&bits, 32);
> >> 
> >> This leads to the find helper dereferencing a full unsigned long,
> >> regardless of the size of the actual variable. The unwanted bits
> >> get masked away, but strictly speaking, a read beyond the end of
> >> the target variable happens. Builds under -Warray-bounds complain
> >> about this situation, for example:
> >> 
> >> In file included from ./include/linux/bitmap.h:9,
> >>                  from drivers/iommu/intel/iommu.c:17:
> >> drivers/iommu/intel/iommu.c: In function 'domain_context_mapping_one':
> >> ./include/linux/find.h:119:37: error: array subscript 'long unsigned int[0]' is partly outside array bounds of 'int[1]' [-Werror=array-bounds]
> >>   119 |                 unsigned long val = *addr & GENMASK(size - 1, 0);
> >>       |                                     ^~~~~
> >> drivers/iommu/intel/iommu.c:2115:18: note: while referencing 'max_pde'
> >>  2115 |         int pds, max_pde;
> >>       |                  ^~~~~~~
> >> 
> >> Instead, just carefully read the correct variable size, all of which
> >> happens at compile time since small_const_nbits(size) has already
> >> determined that arguments are constant expressions.
> >
> >What is the performance impact?
> 
> There should be none. It's entirely using constant expressions, so all of it gets reduce at compile time to a single path without conditionals. The spot checks I did on the machine code showed no differences either (since I think optimization was doing the masking vis smaller width dereference).
> 
> 
> >
> 
> -- 
> Kees Cook