First NLNet Grant Proposal -- First CPU Architecture And Formal Proof of No Spectre bugs -- 2024-12-324
Project Name: Libre-Chip's First CPU Architecture And Formal Proof of No Spectre bugs
Website/Wiki: https://libre-chip.org/
Abstract
Modern CPUs suffer from a constant stream of new speculative-execution security flaws (Spectre-style bugs). Because of that, software has to constantly add more convoluted and slow mitigations to properly enforce security boundaries, such as a WebAssembly VM needing to add code to prevent leaking data from outside the VM to whatever programs are running inside the VM, or a web browser needing to stop JavaScript code from seeing memory that it should not have access to.
To address this major category of CPU flaws, Jacob Lifshay formulated a way to formally prove that a CPU with speculative execution doesn't suffer from any speculative-execution data leaks -- this should work for nearly any kind of CPU design.
We are working towards building a high-performance superscalar CPU with speculative execution and working on proving that it doesn't suffer from any speculative-execution data leaks, thereby demonstrating that this major category of CPU flaws can be fixed once and for all without crippling CPU performance.
Relevant Previous Involvement
-
Jacob Lifshay -- Worked on designing PowerISA CPUs with Libre-SOC for 5yr, built a simple OoO Superscalar CPU simulator https://salsa.debian.org/Kazan-team/power-cpu-sim, built a RV32I CPU with VGA output in a few weeks that runs a 2.5D maze game https://github.com/programmerjake/rv32. Also is the main author of the Fayalite HDL library
-
Cesar Strauss - Contributed to the Libre-SOC project, mostly on digital design and formal verification. Presented the talk "An introduction to Formal Verification of Digital Circuits" on FOSDEM 2024 (https://archive.fosdem.org/2024/schedule/event/fosdem-2024-2215-an-introduction-to-formal-verification-of-digital-circuits/).
Requested support
Requested Amount €50000
Cost Explanation
Nearly all of the cost is for labor, with some amount set aside for general project management (e.g. tracking funding, server maintenance, etc.), and for hardware costs such as buying bigger FPGAs so the CPU design will fit.
We're aiming for a rate of somewhere around $69305.60/yr per person which is the minimum salary for small employers for 2025 in Washington State USA (where Jacob Lifshay lives at the time). https://www.lni.wa.gov/forms-publications/f700-207-000.pdf
Estimated Budget:
General Project Management
€2000
For tracking funding, server maintenance, and similar.
Hardware Costs
€3000
For buying bigger FPGAs, CI server hardware improvements, Oscilloscopes (if necessary), and similar.
Improvements to Fayalite
€10000
For adding the code for translation to Coq, adding tooling for generating FPGA bitstreams, as well as other improvements to Fayalite as necessary.
Building enough of the CPU to execute multiple instructions
€10000
This covers getting register renaming and ALU and/or branch instructions working, as well as other portions of the CPU design as necessary. This may or may not include actually fetching those instructions from memory, instead the instructions may be supplied by a testing harness.
Implementing Memory Subsystem
€15000
This covers the rest needed for the CPU to be able to fetch instructions from memory and execute them as well as implementing Load/Store instructions.
Work towards the Formal Proof of No Spectre bugs
€10000
This covers working on the Formal Proof of No Spectre bugs as well as improvements to other software/hardware needed for that.
Compare with existing/historical efforts
UPEC, as described in "An Exhaustive Approach to Detecting Transient Execution Side Channels in RTL Designs of Processors"
https://arxiv.org/pdf/2108.01979
According to https://arxiv.org/pdf/2407.12232, UPEC is limited in that it only works on a specific conservative mechanism (preventing speculative load instructions from executing until all prior branches are resolved). by contrast our formal proof method should work even if the CPU runs speculative loads before resolving branches, with much less stringent conditions on those loads.
"RTL Verification for Secure Speculation Using Contract Shadow Logic"
https://arxiv.org/pdf/2407.12232
Contract Shadow Logic works by assuming that memory is partitioned into secret and non-secret regions, by contrast our formal proof method is based on verifying that no data processed by transient instructions (instructions that were speculatively executed and then canceled) can affect any digital signals that leave the processor.
Technical challenges we expect to solve during the project
We expect to solve 2 technical challenges:
- To create a working OoO Superscalar CPU with Speculative Execution, trying to design it such that it doesn't have any Spectre-style vulnerabilities (mostly by doing a better job of tracking any state whatsoever that has speculative modifications). The CPU probably will not yet support some of the features required for high performance (e.g. a complex cache hierarchy) and/or supporting running an operating system and/or floating-point. The plan is to add those features later if they are not completed as part of this grant.
- To make substantial progress on creating a formal proof that the CPU doesn't have any Spectre-style vulnerabilities. This is complex enough and has enough potential problems we may run into that we may not finish by the end of this grant, and may need additional grants afterward to finish, hence why the goal is making substantial progress instead of having the proof 100% complete and working.
To address this general category of CPU flaws, Jacob Lifshay formulated a way to formally prove that a CPU with speculative execution doesn't suffer from any speculative-execution data leaks:
- Assume that to leak data, you must be able to observe some change in the digital signals going from/to the CPU core that depends on the results (defined to include any side-effects that change any state, such as cache state changes or branch predictor changes) of some transient instructions (instructions that were speculatively executed and then canceled).
- Take the concrete cycle-accurate CPU design, and construct a theoretical equivalent that has the additional feature of looking into the future to determine which instructions will be canceled and preemptively forces their results to be zeros (or any other value that doesn't depend on transient instructions' data; since we're doing math here, we don't have to limit ourselves to physically possible designs that respect time flow). The reasoning is that if all transient instructions' results are always zeros, they can't possibly be leaking any useful data, therefore the theoretical equivalent CPU design doesn't have any speculative-execution data leaks.
- Formally prove that all digital signals from/to the concrete CPU core always exactly match the timing and data in all respects of the theoretical equivalent CPU design. This proves that the concrete CPU core can't have any speculative-execution data leaks.
It is expected that automated formal proof software (like SMT solvers such as Z3) will probably be too slow to be usable for the formal correctness proof, therefore the plan is to use Fayalite to generate a translation of the CPU design to Coq or some other similar language for writing formal proofs. Fayalite is a HDL library written in Rust that we have been developing since around April 2024 that targets the FIRRTL intermediate language (used by the Chisel ecosystem for generating Verilog; firtool, the FIRRTL to Verilog translator, is part of LLVM Circt, so is well maintained).
A WIP high-level design of our CPU: https://libre-chip.org/first_arch/index.html
Ecosystem
We plan on working with the OpenPower Foundation for any custom ISA extensions we may need. We are already working with the FIRRTL specification GitHub repo to resolve problems we encounter, as well as with LLVM Circt, and with the Rust Language.
This project benefits Europeans (as well as everyone else) by providing a CPU with good performance that doesn't suffer from Spectre-style flaws, as well as working on providing a method of formally proving if any particular CPU design doesn't suffer from Spectre-style flaws, hopefully encouraging CPU designers everywhere to formally prove their CPU designs to not have Spectre-style flaws instead of the reactionary fixes that they have been using that only covers known Spectre-style flaws and has a constant stream of failures.