1. Hardware sandboxing (i.e., ring not-0) isn't much better, as seen by Meltdown.
2. Most production-ready UNIXish kernels have had support for running untrusted code (namely BPF bytecode) in the kernel for decades.
3. Do you really trust all the code currently running in ring 0 on your computer? In particular, do you trust the executable loader. which handles complex untrusted input? What's the line between "code" and "not code"?
4. On most desktop machines, there's a single user, and malware being unable to get to ring 0 isn't particularly stymied; it can still exfiltrate your files, stream your webcam, log into your bank, etc. Why is ring 0 more of a concern than untrusted code elsewhere? (In fact, on most Linux desktops, malware can wait until the user runs sudo, inject itself in, and then run insmod and get to ring 0 directly....)
5. These same machines make a practice of running untrusted, JITted JavaScript and WebAssemy all the time inside the same sandbox you think is too dangerous, and the sandbox works. Not perfectly, of course, but also certainly much better than, say, the Linux kernel protects itself from local privilege escalations. Why is the same software sandbox too dangerous for use in kernelspace?
Meltdown was a single Intel bug, it did not occur on other CPU architectures or on AMD chips. It was a result of asynchronous permission checking and it is a side-channel disclosure (a non-write bug). It is objectively not as bad as the tens of thousands of buffer overruns and memory write vulnerabilities in software.
> 2. Most production-ready UNIXish kernels have had support for running untrusted code (namely BPF bytecode) in the kernel for decades.
Actually it is a security vulnerability as well. In fact, the Project Zero proof of concept for Variant 1 of Spectre was an attack on the BPF interpreter, not even a JIT, it's even worse with a JIT.
> 3. Do you really trust all the code currently running in ring 0 on your computer? In particular, do you trust the executable loader. which handles complex untrusted input? What's the line between "code" and "not code"?
There are levels of trust, of course. I trust the Linux kernel a heck of a lot more than, e.g. V8. And I work on V8. On the WebAssembly implementation. I didn't want to mention it, but yeah, no, I would not put my own code into the kernel.
> 4. On most desktop machines, there's a single user
Again, levels of trust. I would not, e.g. run most userspace software in the kernel, just because it's so broken it will probably bring down the system. All the other things you mention are made easier, not harder by running in the kernel.
> 5. These same machines make a practice of running untrusted, JITted JavaScript and WebAssemy all the time inside the same sandbox you think is too dangerous, and the sandbox works.
I don't want to scare you, but please don't labor under the assumption that web browsers are 100% secure. We have tons of bugs. I mentioned above that the WebAssembly implementation in Chrome is a lot of my work. We've had security vulnerabilities.
> Not perfectly, of course, but also certainly much better than, say, the Linux kernel protects itself from local privilege escalations. Why is the same software sandbox too dangerous for use in kernelspace?
Objectively, no, it isn't better than the Linux kernel. And yes, it is too dangerous. This is based on the hundreds of security bugs that I've been involved with while working on Chrome, and the hundreds more that I wasn't involved with, and the probably hundreds more that are hiding in there. Yes, we take security seriously, and we are very sober about this.
I still think that this is a case of knowing how the sausage is made. I've operated public-facing Linux systems for many years (but I am not either a kernel nor V8 developer, so you know what you're talking about more than I do) and the Linux kernel is ... not good. You take security seriously, and find hundreds of bugs, and it scares you, which is great. The Linux kernel does not (remember the whole "security bugs are just normal bugs" thing, plus the resistance to architectural improvements that kill bug classes, which as far as I can tell you folks seem to be very excited about).
I would hope that you think that V8 + the associated Chrome sandbox (which, to be fair, I think does not have an equivalent in this project) is secure enough to be exposed to random JavaScript / WebAssembly from malicious parties on the internet running and updating 24/7, and keep things reasonably safe, because a billion people do exactly that. I'm not saying it's perfect or unbreakable - I'm just saying I definitely don't trust Linux to be secure against random userspace from malicious parties on the internet running and updating 24/7.
I'm not saying there may not be flaws in WebAssembly or what Cervus delivers.
Again, do not run untrusted code at ring 0, regardless of sandboxing technology.
So as you said a bug in the system would be a kernel exploit. How would this be any different than a exploit in today's kernels? The result would be the same, a bug fix to a kernel, or a fix to the transpiler.
My point is while user running programs in ring 3 protect from bugs and exploits from trashing the system, but they do not protect from bugs in the kernel from trashing the system. In this case you have semantically moved where the bugs can be from the kernel to compile time.
WEBASM engine bug would == kernel bug; if that system is working right then there should be no issue with running in ring 0. Think of it as compile time protection vs run time protection.
I have just woke up and feel I am doing a terrible job of conveying this.
(side note, is the ring system not just another sandboxing technology just facilitated by the CPU? And have we not just had major failures in some of those systems?)
Because now you moved a massive amount of software (namely, the engine implementation, which includes a dynamic compiler, memory management, runtime system, etc--850,000 lines of code for V8) into the kernel, and you just eschewed the simplest of hardware mechanisms (which have been very carefully designed and tests for 50 years, plus verified and proved formally correct by hardware designers) for a very complex set of software checks that are part of a rapidly changing software system that has had dozens upon dozens of security bugs.
The whole point of defense is depth is to add additional layers of security. E.g in a browser, if software checks fail in, (ring 3 userspace) the sandboxing of system calls still doesn't allow a rogue process to even access the filesystem or make kernel calls. Then, on top of that, hardware address translation means that a compromised process cannot attack other processes. If it's all in one giant address space in ring 0, that means a single vulnerability compromises the entire system.
> My point is while user running programs in ring 3 protect from bugs and exploits from trashing the system, but they do not protect from bugs in the kernel from trashing the system.
The whole point is to reduce the TCB (trusted computing base). Bugs in the kernel are rarer because it's smaller, tested more thoroughly, has a clearer and simpler contract, changes slower, and is written by a smaller set of experts than, e.g. random userspace software.
> I have just woke up and feel I am doing a terrible job of conveying this.
No worries. Here's some background that might be useful for the discussion: https://en.wikipedia.org/wiki/Trusted_computing_base
In general, you want to minimize the trusted computing base (i.e. that running in ring 0), and you don't typically want to put a Turing machine inside it!
I would wager to say a implementation in rust is far safer than all of the code that goes into today's linux kernel to run non wasm binaries.
You linked me to TCB, the entire point of the the OPs link to nebullit and cervus is that with this new strategy a entire new security module and way to think about and run un-trusted code has come about. So linking to old ideals and papers explaining how computer security works today only can be used as how things are done now. When the talk about running in ring 0 comes up they are challenging those very ideas. And doing a good job of showing how it can safely be done.
Baring bugs -- yes bugs can happen -- in both the old way and in the new ideas being tossed around with wasm they both are able to provide a level of security. One relies on hardware that can't be easily changed (microcode, new cpu...) , is hard to audit -- and in some cases impossible to audit.
The new notion of compile time checking with wasm allows for a clean approach to ensuring bad programs don't crash the system. Because it does not rely on hardware, but code it can be updated and audited.
I am not arguing that somebodies show HN fun toy is going to be better than 50 years of progress. But I am arguing that a few good years of investment can jump us past those 50 years into a new age of computing, not bogged down by legacy cpu architectures.
Otherwise neat.