My Reading: Memory Defense

Stack Bounds Protection with Low Fat Pointers

Source Code:

Summary: The research work is an extension of their another work (Heap bounds protection with low-fat pointers). The concept of low-fat pointers are originated in the previous paper, they provide well-details of that too. The basic concept of low-fat pointers is: use the pointer memory itself to calculate its object boundary instead of extending the memory or creating a shadow memory. Additionally split up the memory region into different virtual pages based on object size for performance improvement. However, handling heap memory explicitly is easier than the program maintained stack memory for which they propose a stack allocator which will handle all stack memory operation.

Design: According to the design, when a pointer loads the object information, it has to access a lookup table and use its pointer address divided by 32GB (size of the virtual page) as the lookup to figure out its object size. Next, using the roundup technique with pointer address and object size, it can calculate the base address of object memory. Now, as it has both base address and objects size, it can validate the object boundary, every time the pointer access a different location on that object. The bound check operation will operate over instrumented binary. For heap allocation (e.g. malloc), they can modify the allocator to allocate the object into a variety virtual page. But for stack objects, each of them allocates to the same virtual page dedicated to stack. Due to this, their physical memory is non-fragmented, this improves their memory management. For stack objects, it is important to handle memory efficiently. To cope stack objects to the same design for heap and the same time, don’t hurt the regular stack operation, is critical. Moreover, it is important to be efficient. To calculate lookup table, they use lzcnt instruction, for aligning, they do a masking operation (another lookup table), and they allocate the objects in splited 32GB virtual pages similar to the heap, but now separate each 32GB page for heap and stack. Everything is fine unless the physical memory is now more fragmented due to redesigned stack allocator. So, they propose a memory alias to map every virtual stack page to the same physical memory. They do apply different optimization where to check bounds and where not, but they are not a major contribution.

Advantage: Low fat pointer is memory efficient, but when it comes to performance, the design still has to sacrifice little more memory than it has to.

Disadvantage: First of all, now it is easier to predict where an object is located in the physical memory by only knowing the object size. Although pointer integrity is not a part of the bound check, now it can cause greater disaster if not checked. Memory aliasing design for virtual stack pages are not reliable, the design is obscure.

Discussion: The design and evaluation do not show any hope to the practical implementation of low-fat pointers for stack object. Overall, I realize, low-fat pointers cannot precede shadow memory design.