Prossimo – Memory safety for the Internet's most critical infrastructure
memorysafety.org
memorysafety.org
- A BGP server.
- A DNS/DNSSEC server.
- A WiFi/home network gateway.
Those are critical infrastructure. They're self-contained units which talk well-defined protocols. Their functions don't change much. The first two aren't that price-sensitive. They should be verified down to the bare metal.
Apply to DARPA or NIST or DHS for funding for such a project. You might get it.
We are already working on plans for DNS. BGP has been in the back of our minds for a while but we haven’t had the bandwidth to start seriously looking into that yet. If you have ideas get in touch!
For BGP and DNS, probably. But a few big buyers pushing on that might get it.
An all-Rust home router with WiFi, cable, fiber. and DSL would be useful. The trouble is, if it's written by open source people, it will have zillions of versions and be configured by editing a text file.
Rust, on the other hand, has all the downsides of static compilation combined with the library bloat we see in nodejs.
For a home router to run Rust, you'd probably need to compile the entire router into a single firmware image doing everything or you'd need to build an extensive dynamic linking system on top of the existing Rust infrastructure. Or, alternatively, convince the router manufacturers to spend another dollar or two on their EEPROM, but good luck getting them to do that.
Finally, the core parts of the system (the drivers running the DSL/cable/fiber/wifi hardware) are all closed source blobs that the manufacturers (be it Broadcom, Intel, Cisco, or Qualcom) need you to load and execute, usually written in a language like C with none of the benefits of Rust. Unless you want to rewrite the drivers in Rust (and good luck to you reverse engineering all of those and fighting the cease and desist letters you receive), the core parts of the system are still vulnerable as all hell.
The configuration can be made reasonably simple; OpenWRT has a GUI that can easily be dumbed down enough so normal folk can use it. Even in consumer routers, a lot of configuration is still dumped to text files already, anyway.
https://developer.cisco.com/docs/nx-os/#!languages-and-envir...
- DNS/DNSSEC server is a centralized distributed contradiction
- WiFi/home network gateway is _where_ we fix these problems
It cannot be that they do all the marketing sales speech about about Azure Sphere is the future of security in IoT devices, and then offer a C only SDK as option.
They even went to the effort to try to turn this into a positive thing, when the obvious questions started raining on them, given the marketing sales pitch.
https://techcommunity.microsoft.com/t5/internet-of-things/de...
Languages with non-trivial runtimes like Go or Haskell just don't support that same model. You would want to convert your libraries in a certain order so there is one continuous C vs non-C boundary.
Here is the answer; When you wish to provide a library that you intend for use in diverse applications[1] -- and OpenSSL is a excellent example of exactly this case -- you cannot choose a language with a complex runtime. Go has a complex runtime. There is no getting around this; any non-trivial software implemented in Go will require a number of runtime facilities and accommodating that becomes a burden for the library user.
Rust has about the same runtime burden as C/C++. It links the same way as C/C++ and imposes the minimum possible burden on the library user.
[1] diverse as in essentially all platforms, from microcontrollers to high performance systems at scale, over decades of evolution.
They have more information on their website about how to give them money than they do about who they actually are. This makes it easier to route cash to rewrite-it-in-Rust efforts without having to defend these decisions.
Your existing bash scripts don't work in an alternate universe where computers come with ripgrep not grep, but that universe isn't very different otherwise. Some things are a little easier to do, some things are a little harder to do, there are different bugs than in our universe, but it's way less strange than a universe where Lisp Machines still dominate, or even one where the Network uses X.25 rather than IP.
That Google is the man behind the curtain here also comes off as weird since they have not, to my knowledge, standardized their own critical infrastructure on Rust.
That is reasonable. Rust is not the only safe option.
Re everything else....Rust should be seen as the beginning not the end. All those model checking and other "post-hoc" analyses come with major productivity pitfalls. (Wrestling with some solver that may or may not be happy for spooky reasons is profoundly disheartening compared to wrestling with something elegant and compositional like a type system.
Just as there is the ML family of languages and then SML with CakeML, Rust should continue being a fast-moving target where the occasional mistake might happen, but there is academic work revolving around the language which should create a CompCert/CakeML with time.
Some aviation/auto software regulation is a legit, but frankly a some of it doubles as regularity ossification the industry secretly like as it creates a mote around existing businesses that have "evolved" compliance from the get-go. In that case, unproductivity is a feature, not a bug.
I would not hold it up as a realistic model for these sorts of projects to aspire to. Without drastic changes to the way we found unprofitable FOSS foundation infrastructure, whenever productivity and security are at odds, the former will win over the latter.
This argument sounds like rust for rust’s sake.
The rewrite is a 1-time cost.
> Rust’s type system is elegant, but Agda, OCaml F#, F* and Haskell, which are in a whole different league of sophistication, are not?
I am primarily a Haskell user, and I already said I have nothing against GC. Those other languages (except for F#) are great!
> has no specifications or alternative implementations
That is a major downside (I work on GHC and wish it wasn't the only game in town) and affects all of those. That's why I put out SML and the CakeML implementation. SML lacks shiny new stuff but the grunt work has been done.
I would certainly use SML over any variant of C for anything that needs dynamic allocation.
All compilers have miscompilation bugs. Even ones that are proven correct.
> has no specifications or alternative implementations, right?
There is one alternative implementation that is good enough to bitwise reproduce rustc, there are a few other in-progress alternate implementations.
> I’m not aware of any progress in “sealed rust” or “rust belt”
That doesn't mean they haven't progressed. By the way, "Sealed Rust" is now called "Ferrocene" and is expected to GA in late 2022. https://ferrous-systems.com/ferrocene/
> There are other verified compilers that could be used now, though.
This is true, for sure. A compiler being verified is not a panacea.
Unless the exercise is a pretext to rationalize the grant money (which would be the real purpose).
Rust and formal verification work really well together. You can formally verify properties in a small section of critical code, then use the type and lifetime systems to expand those properties to the whole codebase. It's not an either/or, it's a both/and.
Given that, as you correctly point out there are still lots of options that could be used; for example, Ada/SPARK has been around and freely available via GNAT for a long time.
For example, here is a DNS server implemented in Ada/SPARK: http://ironsides.martincarlisle.com/
Latest download is 5 years old - can you find a single instance of IRONSIDES running in the wild?
IMHO the main advantage that Rust has over other approaches that would technically work to create sufficiently secure software is that it has much lower barriers to entry and a much larger community - the hardest part of getting a technology adopted is marketing, and Rust is currently the only option that can compete with C/C++ here.
There is already some promising work on formalizing the semantics of Rust, see for example https://plv.mpi-sws.org/rustbelt/popl18/ and https://people.mpi-sws.org/~jung/phd/thesis-screen.pdf
All these issues I've encountered, or heard of, while working on MISRA-compliant code which is expected to run in cars for years without updates. We make do with C, and extensive tests + ASAN/UBSAN/TSAN can take us a long way, but my point is that Rust is a nice incremental step compared to C. I'm not excluding F* from the list of nice incremental steps, and if the developer ecosystem is there maybe that is a compelling option.
My other point is that MISRA C helps enforce standards in some of the nasty codebases that run in cars, and trust me they can be scary, but it is only a band-aid and I believe that Rust is generations ahead of it.
> but auto industry standards are not reasonable or a language with a garbage collector is not reasonable.
We avoided garbage collection because the run-time is harder to predict, and at the same time Java/Android is used a lot in the infotainment units which are now also running some safety-critical code - with a back-up computer thank you standards - and it makes it harder to allocate resources on the embedded system accordingly.
I will admit I'm all for Rust and will likely try to seek out jobs which I get to use it in.
So, where do I get the version of curl or similar that in fact provides "some of these guarantees" ? Because ISRG isn't spending money on developing software for the Ford Bronco or US aircraft carriers, it is spending money on curl.
MISRA C like a long list of other projects for C or C++ including Bjarne's "Core Guidelines", goes like this:
1. Classify things people will end up needing to do as sinful.
2. Offer them a mechanism to obtain an indulgence, knowing this mechanism will quickly be exhausted.
3. When inevitably this results in tragedy point back at item (1) which absolves you of responsibility.
Example #1, MISRA says you mustn't have any unreachable code. Of course deciding whether code is actually reachable requires solving the Halting Problem, so your automated tooling doesn't do that and will either just flag code that might or might not really be unreachable (useless noise that reduces development velocity) or ignore MIRSA here and just leave your programmers to try not to do this, the status quo ante.
MISRA often needs to dish out "permits" to violate their guidelines because, in practice, the guidelines are too strict to be usable and somebody's workplace requires the guidelines so... the alternative is incredibly creaky and likely unsafe code that's "compliant" - or a "permit" to just ignore a guideline. This is most common when the language doesn't provide a sane way to do something.
Example #2, MISRA forbids hiding variables. If the inner scope has a variable named 'counter' and so does the outer scope, programmers are likely to get confused and mean one when the other is actually in scope, the language compiler won't stop this. But MISRA has to issue a permit because C macros routinely do this and so if your static analysis tool looks at the code it sees lots of such re-use that was (hopefully) harmless.
But if you have hygienic macros (Rust declarative macros) then this isn't a problem, the hygienic macro can't accidentally shadow variables. So you don't need these acrobatics in Rust, if you want to ban shadowing, you can do that and not have weird knock-on effects in macros.
Example #3 MISRA forbids invariants from appearing as conditionals. MISRA reasons that you couldn't possibly have meant to check that buffer != NULL because you know you set buffer in the constructor, and so this is a coding error. But of course you wrote that because you're worried that a bug elsewhere might null out the buffer and blow up the whole system... So now your static analysis tool is screaming that your defensive coding violate MISRA. Do you go ask the Boss for authority to ignore this rule? Or do you rip out the defensive checks?
Tools like F* and WUFFS are different because they've intentionally set their sights low. If you say you want to write the software for your Jet Fighter in WUFFS its authors will carefully explain that this can't be done, and they don't want to help you try. They solve small, well-specified problems in a constrained problem space. They do it very well, you should use them when they're appropriate. They can't be used to replace curl.
None of this would be happening if the market didn't care. I was one of the people who started the Linux kernel modules in Rust project (I'm credited in the announcement), but I started it because I thought it would be fun, and I have a day job that isn't related to Rust at all. So I've spent almost no time on it recently.
The corporate sponsors behind this effort are interested in this because they use Linux (and curl and so forth) at a large enough scale where this sort of thing matters. I would guess that, for instance, Google cares about getting Rust into Android's kernel because they want to credibly claim that Android has fewer zero-days than iOS, and they have a clear public agenda of wanting to get rid of zero-days (see e.g. https://googleprojectzero.blogspot.com/p/0day.html).
So that's their profit motive, and I'm very happy it exists, because this work is too important to rely on the free time of people like me.
(To be abundantly clear: I am not paid by any of the involved organizations, and I had no advance knowledge of this before the public announcement.)
It is going to be very hard to achieve this when iOS is moving into the direction to have 100% of userspace API available to Swift, while Android team forces us to use C and C++ for anything related with real time audio, machine learning and 3D API.
Now in Android 12 even Renderscript was replaced with C and C++ alongside Vulkan compute shaders.
The kernel can even be 100% Rust, it won't help those of us that are cloning Oboe or Vulkan github repo, dealing with JNI boilerplate, C and C++ toolchain issues on NDK.
But, yeah, a company whose business model is tricking old people into buying junk and surveilling their users probably shouldn’t be implicitly trusted.
That’s distracting as heck and I can’t think of a good reason to have such a weird highlight…
The Linux kernel project sounds interesting. It would be nice if there was an effort to purge as much C code as possible from the core of the kernel and really enforce standards such that even a maliciously designed kernel module should not be vulnerable to memory corruption bugs.
May 26, 2021 CWE-416: Use After Free
May 26, 2021 CWE-457: Use of Uninitialized Variable
Putting Rust into kernel seems to me that is a lot of effor for microscopic gains.
Although it would be great to eventually see it.
Linux is thirty years old. And Linux isn't something from whole cloth, it's a Unix, it didn't have some legal and technical baggage that UNIX® brand Unix had but it's a Unix. And Unix is over sixty years old.
We may very well be done. Unix isn't the best conceivable Operating System, but it seems like it's good enough, and in being good enough it became so enormously popular that any new alternative has to be night-and-day better or why bother?
I'm reminded of the Two's Complement Integer thing. Historically it was unclear if this was the "best" way for integers to work. Hardware existed, and sold well decades ago, with some other ideas for how integers could be represented and calculated. But, gradually, over time, even if Two's Complement wasn't clearly the best, it was clearly the most popular, and the others weren't clearly better so why fight it? Sure enough, "Two's Complement Integers" get to just be how integers work by default in newer languages.
If Linus and all the subsystems people were screaming "Nothing but C, under any circumstances" then perhaps you could imagine a competitor to Linux comes along in ten, twenty, thirty years. But so long as they remain open to changes that actually have good results, they're going to stay on top.
Considering Google's funding of these ISRG projects, I hope this changes.
On aspect of the Rust-on-Linux worked I've helped out with is trying to make sure standard library abstractions aren't just suitable for userspace. I hope the medium term result of this is more comparability between bare metal projects on axes where they aren't trying to differentiate.
Longer term, I hope that means a corpus of driver that can be shared between kernels, revolutionizing kernel diversity the same way LLVM revolutionized programming language diversity.
But likely you were thinking of mod_ssl. Except, mod_ssl is how Apache ends up using OpenSSL, where the famous Heartbleed bug is literally exactly the sort of memory unsafety that you'd never do in Rust.