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

[Kees Cook]

On Tue, Dec 07, 2021 at 03:39:30PM -0800, Yury Norov wrote:
> Bitmap functions work identically for all sizes from 0 to INT_MAX - 1. 
> Users don't 'choose between functions based on the size of their target'.
> [...]
> for_each_*_bit() requires a pointer to an array of unsigned longs. If
> it's provided with something else, this is an error on a caller side.

I have a sense we're both talking past each other. Let me try to show
what I'm seeing: the code in the bitmap API chooses one of two paths:

- if "size" is a constant expression and is sizeof(unsigned long) or
  smaller, mask to "size" and call into ff*() built-ins.
- else, use a dynamic sized loop against a series of unsigned longs.

For the dynamic size case, yes, absolutely, things must stay unsigned
long aligned, etc, that makes sense. I don't want to change anything
there.

For the constant-expression size, this requirement does not hold. Every
helper performs a constant expression sized mask of the argument before
using ff*(), for example:

                unsigned long val;

                if (unlikely(offset >= size))
                        return size;

                val = *addr & GENMASK(size - 1, offset);
                return val ? __ffs(val) : size;

In this case, the argument does _not_ need to be a pointer to native
unsigned long. And this is seen in the many many places in the kernel
using non-unsigned-long arguments with a constant-expression size. (e.g.
an int with size 32).

-- 
Kees Cook