Log in

Previous Entry | Next Entry

It has been more than two years since I posted my last verification challenge, so it is only natural to ask what, if anything, has happened in the meantime. The answer is “Quite a bit!”

I had the privilege of attending The Royal Society Verified Trustworthy Software Systems Meeting, where I was on a panel on “Verification in Industry”. I also attended the follow-on Verified Trustworthy Software Systems Specialist Meeting, where I presented on formal verification and RCU. There were many interesting presentations (see slides 9-12 of this presentation), the most memorable being a grand challenge to apply formal verification to machine-learning systems. If you think that challenge is not all that grand, you should watch this video, which provides an entertaining demonstration of a few of the difficulties.

Closer to home, in the past year there have been three successful applications of automated formal-verification tools to Linux-kernel RCU:

  1. Lihao Liang applied the C Bounded Model Checker (CBMC) to Tree RCU (draft paper). This was a bit of a tour de force, converting Linux-kernel C code (with a bit of manual preprocessing) to a logic expression, then invoking a SAT solver on that expression. The expression's variables correspond to the inputs to the code, the possible multiprocessor scheduling decisions, and the possible memory-model reorderings. The expression evaluates to true if there exists an execution that triggers an assertion. The largest expression had 90 million boolean variables, 450 million clauses, occupied some tens of gigabytes of memory, and, stunningly, was solved with less than 80 hours of CPU time. Pretty amazing considering that SAT is NP complete and that two to the ninety millionth power is an excessively large number!!!
  2. Michalis Kokologiannakis applied Nidhugg to Tree RCU (draft paper). Nidhugg might not make quite as macho an attack on an NP-complete problem as does CBMC, but there is some reason to believe that it can handle larger chunks of the Linux-kernel RCU code.
  3. Lance Roy applied CBMC to Classic SRCU. Interestingly enough, this verification could be carried out completely automatically, without manual preprocessing. This approach is therefore available in my -rcu tree (git://git.kernel.org/pub/scm/linux/kernel/git/paulmck/linux-rcu.git) on the branch formal.2017.06.07a.

In all three cases, the tools verified portions of RCU and SRCU as correct, and in all three cases, the tools successfully located injected bugs. (Hey, any of us could write a program that consisted of “printf(“Validated/n");”, so you do have to validate the verifier!) And yes, I have given these guys trouble about the fact that their tools didn't find any bugs that I didn't already know about, but these results are nevertheless extremely impressive. Had you told me ten years ago that this would happen, I have no idea how I would have responded, but I most certainly would not have believed you.

In theory, Nidhugg is more scalable but less thorough than CBMC. In practice, it is too early to tell.

So what is the status of the first five verification challenges?

  1. rcu_preempt_offline_tasks(): Still open. That said, Michalis found the infamous RCU bug at Linux-kernel commit 281d150c5f88 and further showed that my analysis of the bug was incorrect, though my fixes did actually fix the bug. So this challenge is still open, but the tools have proven their ability to diagnose rather ornate concurrency bugs.
  2. RCU NO_HZ_FULL_SYSIDLE: Still open. Perhaps less pressing given that it will soon be removed from the kernel, but the challenge still stands!
  3. Apply CBMC to something: This is an ongoing challenge to developers to give CBMC a try on concurrent code. And why not also try Nidhugg?
  4. Tiny RCU: This was a self-directed challenge was “born surmounted”.
  5. Uses of RCU: Lihao, Michalis, and Lance verified some simple RCU uses as part of their work, but this is an ongoing challenge. If you are a formal-verification researcher and really want to prove your tool's mettle, take on the Linux kernel's dcache subsystem!

But enough about the past! Given the progress over the past two years, a new verification challenge is clearly needed!

And this sixth challenge is available on 20 branches whose names start with “Mutation” at https://github.com/paulmckrcu/linux.git. Some of these branches are harmless transformations, but others inject bugs. These bugs range from deterministic failures to concurrent data races to forward-progress failures. Can your tool tell which is which?

If you give any of these a try, please let me know how it goes!


( 2 comments — Leave a comment )
Alex Groce
Jun. 8th, 2017 08:22 am (UTC)

Are these "our" mutants, or is there some other source of the Mutation branches?
Jun. 9th, 2017 05:34 pm (UTC)
Re: Mutants?
At least one is a mutant from our past work, and several are higher-level mutants that we discussed for future work. I expect that it would be good to include more of the past mutants, but I have been a bit slow about getting back through the spreadsheets.
( 2 comments — Leave a comment )