Society is increasingly reliant on software, but deployed software contains security vulnerabilities and other bugs that can threaten privacy, property and even human lives. When a security vulnerability or critical error is discovered, a software patch is issued to attempt to fix the problem, but patches themselves can be incorrect, inadequate, and break necessarily functionality. This project investigates the full workflow for the developer to rapidly diagnose the root cause of the vulnerability or error, for the developer to test that a prospective patch indeed completely removes the defect, and for users to check the issued patch on their own configurations and workloads before adopting the patch.
Record/replay is an emerging technique for low-overhead recording of deployed software, in case a failure or exploit occurs in the field, so the user can transmit the recorded log to the developer to deterministically reproduce the bug. But existing record/replay systems can replay only with the exact same version of the software as recorded, possibly with insertion of debugging instrumentation that does not affect application state. This project investigates “mutable replay” to help reproduce, diagnose, and fix software bugs. Mutable replay will enable logs recorded with the buggy version to be replayed after the modest code changes typical of critical patches, which do impact application state, to show that patches work correctly (or not!) to resolve detected problems.
This project leverages semantic information readily available to the developer (e.g., from their version control repository) to conduct static and dynamic analyses to correctly transform the recorded log to enable mutable replay. For example, the replayer could skip over no longer valid log entries and temporarily go-live when new data is needed.
We are currently modifying rr (https://rr-project.org/) to support mutable replay, since rr is an actively supported open-source record/replay system in wide use and integrated with gdb. Its recorded log includes details of the instruction locations in memory, register contents, and other low-level details rather than just the non-deterministic inputs to the application recorded by some other record/replay systems. This makes it impossible to use rr’s replayer to run the new version of the executable while feeding from a log recorded with a previous version of the executable. Instead, we need to patch the old executable to jump into the new executable where they diverge, run that modified code “live”, then jump back if/when they converge.
We would also like to investigate mutable replay for Java/JVM, but first need to find a practical record/replay system to leverage — or build our own.
This project is seeking to admit and fund a new PhD student. We are also recruiting MS and undergraduate students for independent study projects. Students should be interested in compilers and language runtime environments, debuggers, program analysis, software testing, software reliability, security and systems (Linux kernel and JVM).