Memory Safety Instrumentations in Practice: Usability, Performance, and Security Guarantees

Abstract

Memory safety violations due to C's undefined behavior, although well researched, still cause security breaches year by year. The most dangerous reported violations are spatial safety violations, where objects are accessed outside of their bounds. A wide variety of spatial safety sanitizers promise easy usage, broad security guarantees, and a low execution time overhead. However, only few of them are actually used. Instead of proposing yet another sanitizer, we dig deep into Low-Fat Pointers and SoftBound, two approaches to generate fast-to-execute safe programs with strong safety guarantees, and identify pain points in their usage. We found that seemingly small simplifying assumptions or limitations of the approaches often lead to spurious error reports. On top of analyzing usability issues, we set up a framework that abstracts common tasks of memory safety instrumentations, such as finding locations for checks and eliminating redundant checks. This abstraction allows us to draw a fair comparison between approaches when it comes to execution time and the number of safe accesses. We use this framework to give novel insights into how many accesses are provably safe, and where to attribute execution time overhead. Our findings help future research on memory safety instrumentations by identifying issues that current approaches face in their practical application. We make our LLVM-based instrumentation framework available to reduce the effort required to implement new instrumentations and to ease comparisons to Low-Fat Pointers and SoftBound.