Lossless instruction-to-object memory tracing in the Linux kernel

Abstract

The lack of visibility into Linux's behavior makes it hard to refactor and maintain. To peer inside the box, we present Memorizer, a self-contained, low-level tracing framework that tracks (most) object allocations, data accesses, and function calls within the kernel. The core insight is a low-level object-centric representation that records detailed lifetime information while linking each operation (call/read/write) with its intended target. We evaluate Memorizer using extensive input programs and demonstrate its value by showing how Memorizer can (1) aid in refactoring, (2) extend code coverage with object coverage to improve testing and analysis, and (3) identify leaky abstractions. We also release a large data set, visualization tools, and Memorizer's source. This generic, object-centric approach is the first to provide lossless instruction-to-object tracing, adding an essential software engineering capability to the overly complex Linux kernel.