LLVM based Virtual Machine "Environment" idea sanity check.

I've got an idea for a program, and after readig about 1/3 of your documentation, I think LLVM is what I'm looking for.

What I'd like now is some help to see if my idea is "sane" and and shed light and direction that could be provided.

I want to build a simplified "Virtual Machine" containing:

A Terminal
Hard Drives (image files)
Some Kind Of Networking Device

LLVM programs would be run "inside" this Virtual Machine, accessing the terminal and hard drive images and networking device... but not having access to ANYTHING on the host computer (except through the "virtual" devices).

I had originally planned on writing my own "processor core" for this project... but I'd rather use LLVM (Mainly because I don't have to write my own high level tools).

Here's what I'm thinking I need to do. It seems to me that I have to "port" LLVM using the System Library to my "Virtual Machine" (which also includes a bit of magic of the "Exokernel" operating system stuff). Then I can run LLVM programs in my "port" on my "environment" to get what I want.

Then LLVM can JIT compile programs but still only have access to my "Virtual Machine." I think.

Any pointers on where to read, Ideas to move on... would be much appreciated.

Thanks.

Hi Shawn,

The idea reminds me a lot on the research on Java operating systems. If
you google for it, you'll find a plenty of references.

In my opinion, the main problem you're going to run into if you want to
design a serious system is interprocess communication. With Java, that's
a huge problem, because each app traditionally runs in a separate virtual
machine. Garbage collection complicates this even further.

You might also want to check Singularity project:

Shawn "AutoDMC" Boles wrote:

I've got an idea for a program, and after readig about 1/3 of your documentation, I think LLVM is what I'm looking for.

What I'd like now is some help to see if my idea is "sane" and and shed light and direction that could be provided.

I want to build a simplified "Virtual Machine" containing:

A Terminal
Hard Drives (image files)
Some Kind Of Networking Device

LLVM programs would be run "inside" this Virtual Machine, accessing the terminal and hard drive images and networking device... but not having access to ANYTHING on the host computer (except through the "virtual" devices).
  

If you don't mind my asking, can you tell us a little more about your overall goal for this project? How stringent is your isolation requirement? What types of programs do you want to run on this VM (programs written in a special language, small C applications, a full operating system like Xen/VMWare, etc)? LLVM can probably make development go faster, but the sanity of your project greatly depends on what it is for and just how much it will do.

I had originally planned on writing my own "processor core" for this project... but I'd rather use LLVM (Mainly because I don't have to write my own high level tools).
  

I think LLVM would probably make a good "processor core" for the very reasons you mention.

Here's what I'm thinking I need to do. It seems to me that I have to "port" LLVM using the System Library to my "Virtual Machine" (which also includes a bit of magic of the "Exokernel" operating system stuff). Then I can run LLVM programs in my "port" on my "environment" to get what I want.
  

This doesn't make sense to me. If you're going to build your virtual machine out of LLVM components, why would you make those components run on the virtual machine itself?

Then LLVM can JIT compile programs but still only have access to my "Virtual Machine." I think.
  

The difficulty of this depends on the scope of your virtual machine. Unlike other bytecode languages, LLVM can represent programs with buffer overruns and type unsafe casts, both of which could undermine any restrictions that your VM places on programs that it runs.

Now, you could use LLVM to build a safe virtual machine, but you would have to figure out how to solve these problems. There are various solutions of varying degrees of sanity, but I can't really give any feedback on which is more appropriate until I know more about the constraints of your project.

-- John T.

If you don't mind my asking, can you tell us a little more about your overall goal for this project?

> Snip
> I can't really give any

feedback on which is more appropriate until I know more about the constraints of your project.

-- John T.

Hopefully I can explain my project more fully (without being too wordy):

"CRAZY CRACKPOT IDEA"

What I want to do is create an idealized "processing node." This virtual machine would include:

1) A processing core
2) Access to "permanent storage"
3) Access to a "Network device"

With any luck, you should be able to run this virtual "processing node" on an old Pentium II, on a Apple G3, on a soon-to-be-extinct PS2 (because everyone'll be selling them to get the PS3), on a Gumstix Waysmall computer... basically anything that you can get your hands on.

Each of the computers above would run the "idealized processing node" in their native operating system.

An outside observer would see a group of heterogeneous computers networked with each other through the internet.

An "inside" observer would see a group of homogeneous "processing nodes" networked with each other through Jabber (see below).

"WHAT I WANT TO USE"

I'd like to use LLVM for the Processing Core part of this "idealized processing node." This would allow me to exploit the programming language frontends available and the optimizing and JIT compiling abilities on the backend. The key here is that the "node operator" might want to help a project, but might not want to fully trust the code for the project. For example, if I run Folding@home, I have to fully trust the programmers of the Folding@home client... trust that they haven't hidden something evil in their program, or have a hidden vulnerability that hurts me.

However, if the Folding@home process was running in a virtual machine, they could run just about any code they wanted in it, without me having to worry (in general). Of course, this is where JIT compiling should help out... Basically, the process should have access to the computing abilities of the host... but no access to any of the hosts actual hardware devices. As far as this hypothetical Folding@home process is concerned... it's running directly on a simple LLVM processor with a hard drive and a NIC... that's it.

Also, the Folding@home team has to be able to handle multiple operating system's way of handling files and permissions. My "processing node" would give them a complete "raw hard drive" that they could write to, without worrying about permissions. They could use whatever data storage scheme they found useful, and I would know that no matter what they wrote, they couldn't clobber any of MY data. Worst case, their hard drive image could fill my physical hard drive... but that can be remedied in the somewhat harsh but forceful way of killing the offending process and deleting their hard drive image.

As for Networking... I was thinking of using something high level, like Jabber. While Jabber was designed for instant messenging... it could easily be used for "interprocess communication." The hypothetical Folding@home client would contact "ProjectManager@foldingathome.org" to get it's data. The processing node's "Jabber ID" would be something like "node@domain/foldingathome". The server could send updates to the node, and the node could send updates to the server... all without worrying about "IP addresses."

The general all around idea *from the user's side* is to make a way for a node to be able to run trusted/untrusted code as fast as possible in as safe a way as possible.

The general all around idea *from the project developer's side* is to have a system that allows the developer to write one version of a client and run it on as many "virtual processing nodes" as possible.

The general all around idea *from the point of view of the program being run* is that it's the only process running on top of very simple "hardware"... a LLVM processor, a hard drive, and a NIC that speaks Jabber.

My goal is to write the blank-slate VM that is as simple and powerful as possible... and let the application writers string the "processing nodes" up how they want. If they want a more complex Interprocess Communication... they can write it on top of the plain-text-through-Jabber protocol (maybe with XML or YAML or something even more exciting). Or they can write a library that makes a full UNIX style file system on top of my raw-block "hard drive." They can make the nodes run single file or in a tree or any way they want... I'm just building the "hardware platform."

As for the "processing node" itself... the user can kill it, or restart it if it seems to have become "hung..." or the server requesting the processing job could ask to have it's job reset or killed. The most important thing is that a badly written program can "nuke" and "crash" the VM, but it shouldn't "nuke" or "crash" the host computer.

This is where I'm stumped... how can I use LLVM inside my "simplified processing node," allowing it only access to my "virtual devices," getting it to run "as fast as possible" with as little possibility of it clobbering the host computer should something untoward happen to the LLVM code? I'm still sifting through the documentation and feel like I'm 90% there to understanding out LLVM works as a Virtual Machine, and not a Compiler backend or Intermediate Language. Or, if LLVM isn't my "perfect match," where should I look?

LLVM isn't a perfect match today. It does provide the ability to transport code around in a portable way and execute it, but lacks these features:

1. Given C input code, the output isn't guaranteed to be portable. This
    is a C limitation, not an LLVM limitation. In practice, if you avoid a
    few constructs, and stay away for complex system interfaces, the code
    you write will be mostly portable.
2. LLVM doesn't provide safety (it doesn't prevent the program from
    clobbering the machine).

#1 is a limitation of C, not LLVM. If you use a portable source language, you'll get portable LLVM Code. Several people are interested in #2 and have solutions with various tradeoffs, if you ask here I'm sure they'll tell you about them.

In practice, the hard part, and the value add, of your project seems to be the higher-level design issues (what APIs / interfaces do you provide, how to pull it together into a useful system, how to set up the middleware, etc). Building the initial system on top of LLVM without worrying about #1/#2 seems like a good way to get a prototype up and running, and you might find out that #1/#2 might get solved for you by other community members.

-Chris

Hi Shawn,

What I want to do is create an idealized "processing node." This
virtual machine would include:

1) A processing core
2) Access to "permanent storage"
3) Access to a "Network device"

With any luck, you should be able to run this virtual "processing node"
on an old Pentium II, on a Apple G3, on a soon-to-be-extinct PS2
(because everyone'll be selling them to get the PS3), on a Gumstix
Waysmall computer... basically anything that you can get your hands on.

Each of the computers above would run the "idealized processing node" in
their native operating system.

Have you looked at the Inferno OS? Originally from Bell Labs, now
developed by Vita Nuova.

    http://www.vitanuova.com/inferno/index.html

It sounds just like what you're thinking of. Even if you do want to
role your own, with or without LLVM, you'll pick up lots of useful
information from the papers on it.

Cheers,

Ralph.