WasmBoxC: Simple, Fast, and VM-Less Sandboxing

The software ecosystem has a lot of useful but unsafe code, and the easier it is to sandbox that code, the more often that’ll happen. If it were as simple as passing the compiler a --sandbox flag that makes an unsafe library unable to see or affect anything outside of it, that would be incredible! We can’t get it quite that easy, but this post describes WasmBoxC , a sandboxing approach that is very simple to use. All you need to do is:

• Compile the unsafe library using a WebAssembly (wasm) compiler instead of the normal system compiler . That uses wasm internally, but you don’t need to care about that — all you see is it emits a C file with sandboxed code.
• Write some C to interface with the compiled C of the unsafe library . (This is necessary because the sandboxed code can’t access outside memory, and also it uses the portable wasm ABI.)

Compile and link that C code, and now that unsafe library is sandboxed from the rest of your application! In a later section we’ll see concrete examples of how easy both those steps are.

Here is the approach in more detail:

By compiling to wasm we sandbox the code, preventing it from accessing anything on the outside. That includes both memory - the sandboxed code can’t read or write to anywhere outside it - and capabilities - the sandboxed code can’t do anything but pure computation, unless you give it a function to call to do things like read from a file, tell the time, etc. We also get the rest of the wasm guarantees on safety and portability . Wasm sandboxing is even safe to run in the same process as other code (at least modulo Spectre-type vulnerabilities), much like Software Fault Isolation (SFI) .

After we’ve compiled an unsafe library to wasm, how can we run it as part of our application? We could integrate a wasm VM and run the wasm there. But instead, with WasmBoxC we take a VM-less approach and compile the wasm into native code, while preserving the wasm semantics, including the sandboxing. That native code can be linked normally into an application, which is much simpler than integrating a wasm VM.

The specific approach WasmBoxC takes to compile wasm into native code is to compile it to C using wabt ’s wasm2c tool, and then run a standard C compiler on it. In fact, WasmBoxC’s approach compiles into a simple subset of C . This is a big part of what makes this approach so simple, and brings several advantages:

• It’s easy to read and verify the generated C for security.
• It lets us use a C compiler like clang or gcc to make the sandboxed code very fast.
• It’s very easy to use the code in existing build systems.
• It’s easy to write code to interact with the sandbox.
• A single build of C code can be compiled and run on practically any platform, and code to interact with the sandbox also only needs to be written once.

Despite the simplicity of using C, WasmBoxC sandboxing has low overhead: just 14% with some non-portable C code (the “signal handler trick”, see later), or 42% in 100% portable C (with no OS- or CPU-specific operations at all). We’ll also see that there are options in between those 14% and 42% numbers.

The basic idea in WasmBoxC is simple and not original . What is new in this post is showing that the approach works, doing benchmarking on real-world code to show it is fast, presenting complete examples of how easy it is to sandbox real-world libraries, and writing up the approach in detail to describe the benefits (see in particular the section on memory-safe languages). This post also invents a name for the technique.

Speed

To get an idea of WasmBoxC’s speed, let’s take a look at 20 benchmarks , comparing clang 9.0.1, clang 11 (dev version as of May 23 2020), gcc 9.2.1, and WasmBoxC. All numbers are normalized to clang 9 (which is therefore equal to 1; lower numbers are better).

These benchmarks include a wide variety of code, and the ones prefixed with zzz_ are real-world codebases or benchmarks: the Box2D and Bullet physics engines, the CoreMark and LINPACK benchmarks, the Lua VM (one GC and one computational benchmark), the LZMA and zlib compression libraries, and the SQLite database. Incidentally, this shows WasmBoxC can run all of these today!

Two results are shown for WasmBoxC, representing two implementations of memory sandboxing. The first is explicit sandboxing, in which each memory load and store is explicitly verified to be within the sandboxed memory using an explicit check (that is, an if statement is done before each memory access). This has 42% overhead.

The OS-based implementation uses the “ signal handler trick ” that wasm VMs use. This technique reserves lots of memory around the valid range and relies on CPU hardware to give us a signal if an access is out of bounds (for more background see section 3.1.4 in Tan, 2017 ). That is fully safe and has the benefit of avoiding explicit bounds checks. It has just 14% overhead! However, it cannot be used everywhere (it needs signals and CPU memory protection, and only works on 64-bit systems).

There are more options in between those 14% and 42% figures. Explicit and OS-based sandboxing preserve wasm semantics perfectly, that is, a trap will happen exactly when a wasm VM would have trapped. If we are willing to relax that (but we may not want to call it wasm if we do) then we can use masking sandboxing instead (see section 3.1.3 in Tan, 2017 ), which is 100% portable like explicit sandboxing and also prevents any accesses outside of the sandbox, and is somewhat faster at 29% overhead. Other sandboxing improvements are possible too - almost no effort has gone into this yet.

An interesting thing happens in the lua_binarytrees and havlak benchmarks, where WasmBoxC is actually faster than both gcc and clang in all sandboxing modes, up to 32%! How can we beat normal native builds, and by so much? Looking into this , both of these benchmarks use a lot of malloc s and data structures with pointers. Like the x32 ABI , wasm is 32-bit, so pointers take half the space. Measuring the maximum process memory used in lua_binarytrees , WasmBoxC uses 33% less which helps a lot with CPU cache usage. While this makes a big difference on these two benchmarks, we are likely getting some speedup on the others as well due to this factor, as on average x32 is faster than normal x64 by around 5-8% . Wasm is a nice way to get something like x32’s benefits!

The benchmarking here measures performance within the sandbox. It does not measure the speed of calls from the outside in or inside out. Such calls can be very fast because the sandboxed code is just C, which means that we can even inline across the sandbox boundary — safely! — if we do LTO. I verified that happens in the sandboxing example in the next section, see later. (Note, however, then using the signal handler trick may make things more complicated here.)

Some final notes on performance:

• We can compile WasmBoxC’s C code with any native compiler. In the above we did so always with clang 9 for simplicity. The results vary a little when changing the compiler, for example the “explicit” sandboxing results go from 14% up to 16% with gcc 9.2 or down to 11% with clang 11. There’s nothing magical about that 14% figure — we are at the point where native compiler differences matter.

• Results should improve over time as wasm adds more performance features like simd (note that the native compilers we compared to may have gained an advantage from autovectorization).

• That WasmBoxC can reach 14% overhead shows that the cost of compiling through wasm (which cannot represent irreducible control flow, for example) is fairly low, and also that current compilers to wasm are not introducing significant unnecessary overhead.

I’ve done my best to measure everything here carefully and accurately, but it’s possible I’ve made a mistake somewhere. Please check my work and see if you get similar results!

Ease of use

This section has a full example of WasmBoxC usage. Here are the source files:

// my-code.c

#include <stdint.h>
#include <stdio.h>

// We could also include the .wasm.h file for these,
// but let's declare externs manually for the example.

extern void wasmbox_init(void);

extern uint32_t (*Z_twiceZ_ii)(uint32_t);

extern uint32_t (*Z_do_bad_thingZ_ii)(uint32_t);

int main() {
  puts("Initializing sandboxed unsafe library");
  wasmbox_init();
  printf("Calling twice on 21 returns %d\n", Z_twiceZ_ii(21));
  puts("Calling something bad now...");
  Z_do_bad_thingZ_ii(1);
  puts("(this will never be printed, as the bad thing will trap)");
}

main() is pretty simple: Initialize, call something in the sandbox that does a computation for us, and call something that will trap inside the sandbox. (What’s with the Z_ stuff? See later in “API”.)

// unsafe-lib.c

#include <stdlib.h>

__attribute__((used))
int twice(int x) {
  return x + x;
}

__attribute__((used))
int do_bad_thing(int size) {
  // Allocate an unknown size here (so the LLVM optimizer doesn't know if the
  // store later down is valid or not).
  char* x = malloc(size);
  // Write to an address that is definitely not in the sandbox (the default
  // memory size is much smaller), which in wasm will trap.
  x[1024 * 1024 * 1024] = 42;
  // Avoid the optimizer knowing the store can never be observed.
  return (int)x;
}

twice() does what you’d expect, and do_bad_thing does a store that will definitely trap. (Ignore the details there; for this example we need the LLVM optimizer not to remove the bad code as undefined behavior!)

And here’s how easy it is to use WasmBoxC to get a fully sandboxed library linked with our normal code:

# build our main code to an object normally
$ clang my-code.c         -c -O3 -o my-code.o
# build the unsafe library to C with emcc
$ emcc unsafe-lib.c          -O3 -o unsafe-lib.wasm -s WASM2C --no-entry
# build the unsafe library's C to an object normally
$ clang unsafe-lib.wasm.c -c -O3 -o unsafe-lib.o
# link normally
$ clang my-code.o unsafe-lib.o -o program

Pretty simple! The only “interesting” part here is that the second command uses a wasm toolchain to compile the library to wasm and then to C. Here we use Emscripten (see the download instructions for how to get it; the zlib example in the next subsection also covers that). Note that there is no -c there, since from emcc ’s point of view it is a full compile+link to wasm, after which it runs wasm2c for us, for which we pass -s WASM2C . Note also that we tell it --no-entry since we are building a library here.

Running ./program (just a normal executable, no VM here!) we get this:

$ ./program
Initializing sandboxed unsafe library
Calling twice on 21 returns 42
Calling something bad now...
[wasm trap 1, halting]

Exactly as expected!

Aside from it being easy to build code with WasmBoxC, it’s also easy to see how it works: just read the generated C code. For example, we called Z_twiceZ_ii earlier. If we’re curious what that is, we can just read the C and see that aside from the function pointer indirection it’s just a plain C function:

static u32 w2c_twice(u32 w2c_p0) {
  // [..the code in the body, ending in a return..]
}

We can see exactly what parameters it takes, what it returns, read the typedef for u32 , and so forth. The actual compiled code in the body (omitted here) is not very readable, but it is still C code. That lets us do things like add convenient printfs for debugging for example. This is also why we said earlier that LTO can inline across the sandbox: the sandboxed code is just more C code. Here is the call to twice and the printing of its result (in LLVM IR, and before LLVM LTO):

define i32 @main() #0 {
  [..]
  %2 = load i32 (i32)*, i32 (i32)** @Z_twiceZ_ii, align 8, !tbaa !1
  %3 = tail call i32 %2(i32 21) #8
  %4 = tail call i32 (i8*, ...) @printf(i8*
    getelementptr inbounds ([32 x i8], [32 x i8]* @.str, i64 0, i64 0), i32 %3)

and here is the result after LLVM LTO:

define i32 @main() #0 {
  [..]
  %6 = tail call i32 (i8*, ...) @printf(i8* nonnull dereferenceable(1)
    getelementptr inbounds ([32 x i8], [32 x i8]* @.str, i64 0, i64 0), i32 42)

Note how before LTO we load a function pointer, then call that with 21 , then printf that result. The function pointer is there because wasm2c emits very flexible code, more than we need in fact. We may want to add an option to avoid that indirection, but as you can see, LTO can fix that up already: in fact it manages to replace the call to twice(21) to a constant 42 in main , avoiding a call entirely! Optimizing across the sandbox boundary can be hugely beneficial, and it’s fully safe.

A final note on the emitted C code: if you do read it you’ll notice it doesn’t look very optimized. That’s because wasm is very low level itself and wasm2c translates it in a simple and accurate way — it doesn’t try to emit “optimal” C code. That simplicity means that to be fast we depend on C compiler optimizations, which is why the example builds with -O3 .

WasmBoxing zlib, and memory management

For a more complete walkthrough of porting an entire real-world library and not just a single file, see this simple gist which shows how easy it is to sandbox the zlib compression library. That includes the full details of how to get the wasm toolchain, so you can follow those instructions step-by-step from scratch.

That also shows examples of how to do memory management in the sandbox, which is very simple:

• The “memory” seen by the sandboxed code is a single buffer of memory that is malloc ed by the runtime. The sandboxing guarantees that the compiled code can only access that buffer, and nothing else.
• Pointers in the sandbox are just 32-bit integers, which refer to locations in that buffer.
• When you get a pointer from the sandbox, you can read that memory directly (that is, there is nothing that prevents the outside from looking in). You can do so by reading from the buffer at the offset of that pointer. That is, if the buffer is at absolute (normal, not in the sandbox) address BUF and you want to read data at the location a sandbox pointer of value ptr refers to, you would read at absolute address BUF + ptr .
• The sandboxed code has its own malloc and free , which look normal to that code, but only reserve and release ranges of memory in the singleton buffer. If you want to pass some data into the sandbox, a simple way is to malloc in the sandbox (using the sandbox’s malloc ) and copy the data in.

Other WebAssembly toolchains

This could be even simpler if we used clang for everything and not emcc. We can do that too! There is nothing specific to Emscripten about the WasmBoxC approach, all we need is a compiler to wasm, wasm2c, and C runtime support for wasm2c’s output. Plain clang could work too since it has wasm support; assuming you already use clang, the only new build tool you’d need to add is wasm2c. Or you could use the WASI SDK or anything else. Adding wasm2c integration and runtime support in Emscripten was not too hard and it probably would be similar elsewhere.

I focused on Emscripten because I’ve already done the work to integrate wasm2c there forother reasons anyhow. Emscripten also supports porting the widest range of software currently which is how we could run all the codebases mentioned in the benchmark section. And Emscripten does a lot of useful optimizations to emit fast wasm which helps see how fast WasmBoxC can be.

Related things

SFI

Software Fault Isolation has already been mentioned in the introduction, with WasmBoxC mainly differing from it in the use of wasm and compiling that through C. As a result, one specific limitation in the WasmBoxC approach is that you must compile the code from source, unlike SFI methods that can operate on a binary.

SFI using masking can achieve overheads of 12% , which is significantly better than the 29% we observed. It is possible that WasmBoxC’s masking sandboxing could be improved — no effort has gone into that yet.

MinSFI

WasmBoxC is similar to MinSFI (itself inspired by asm.js , one of the predecessors of wasm), whose key idea is to convert unsafe code to a sandboxed form. It does so on LLVM IR, while WasmBoxC converts to wasm which is inherently sandboxed (another example in this space is in Kroll, Stewart, and Appel, 2014 which works on CompCert’s IR).

Most of the work in such an approach is to properly define the sandboxed form (without making any security mistakes), and to implement it (which means a lot of tooling work — we need to be able to convert real-world code to that form). When MinSFI was created wasm didn’t yet exist, and there wasn’t a good alternative, but today wasm fits that need very well: all the hard work of speccing and implementing it has been done, and we can just use that in WasmBoxC.

RLBox-wasm

RLBox describes a framework for fine-grain isolation of untrusted code. One of the isolation mechanisms (see section 9 in the paper) uses wasm; call that “RLBox-wasm”.

Like RLBox-wasm, WasmBoxC compiles to wasm and then to native code. The difference is how the native code is generated. RLBox-wasm uses a customized build of Lucet which uses CraneLift to generate native code, while WasmBoxC compiles to C and then uses a standard C compiler like clang or gcc. To consider the performance difference caused by that, let’s look at this quote from the RLBox paper:

We find that Wasm sandboxing [using RLBox with Lucet] imposes a 85% overhead [..] We attribute this slowdown largely to the nascent Wasm toolchains, which don’t yet support performance optimization on par with, say LLVM.

WasmBoxC offers a useful comparison point in this context since it can use LLVM. Indeed, as WasmBoxC has 14%-42% overhead, it supports the quote’s assertion that a large part of RLBox-wasm’s current 85% overhead is due to CraneLift being fairly new (but it is making good progress ). (However, RLBox-wasm is only measured on one benchmark, libGraphite, which limits our ability to generalize.)

Another performance difference between RLBox-wasm and WasmBoxC is that RLBox-wasm has trampolines between the sandbox and the outside. Normally such trampolines do context and stack switching, etc., and can have significant overhead. In RLBox-wasm they use a custom trampoline in Lucet which reduced the overhead by 800%, almost to nothing. In comparison, with WasmBoxC as we saw earlier the sandboxed code is just plain C and there are no trampolines at all, and even inlining across the boundary works.

An advantage RLBox-wasm has is that CraneLift is a dedicated wasm compiler, and so it may use techniques specific to wasm. For example, it might pin a register for the sandboxed memory, or it might use a nonstandard calling convention inside the sandbox. Those things could not be done with plain C code with WasmBoxC. For now it looks like LLVM’s general advantage outweighs using a wasm-specific compiler, but in time that might change.

Another advantage RLBox-wasm has over WasmBoxC is build times: WasmBoxC compiles to C and then runs a full C compiler while CraneLift is designed to compile wasm to native code. The WasmBoxC approach inherently adds extra steps in the compilation process. However, as mentioned throughout this post, going through C has benefits not just to speed but also to ease of use, so overall there are interesting tradeoffs in this space.

edit: It was pointed out to me that an early version of RLBox in January 2019 used wasm2c, which I was not aware of.

Current status

Security

How secure is WasmBoxC? As with anything new you should assume it is experimental for now. However, it is built on well-tested foundations, in particular, wasm itself, wasm toolchain components like clang, and standard C compilers. The key wasm2c component has not been used in production yet to my knowledge, but good indications are that wasm2c passes the wasm spec test suite which covers many sandboxing and portability corner cases, and we’ve fuzzed it .

Another useful thing is that the C output can be inspected for safety. It’s easy to see that all memory accesses go through the same few load/store methods, and that those are guaranteed to stay within the sandbox. While C isn’t a memory-safe language, the very simple form of C we emit should in fact be safe and easily seen as such.

It would be great for security people to take a look at WasmBoxC — please help out!

Compatibility

In terms of what code can be sandboxed right now, in the current implementation it’s basically anything Emscripten can port to wasm, which is quite a lot — it’s used to port many entire game engines , for example, and as we saw it can port all the codebases we benchmarked. In particular it includes support for most of the C and C++ languages and standard libraries, including tricky things like setjmp (which is needed for Lua). However, a few things Emscripten supports have not been enabled in WasmBoxC yet, like C++ exceptions and pthreads.

Other limitations

wasm2c output is a simple subset of C, and it builds with gcc or clang on all machines I’ve tested, but it doesn’t work on MSVC yet (at the moment it uses some gcc/clang intrinsics like __builtin_expect ). Help is welcome in implementing support for that compiler and others!

A concrete limitation right now of WasmBoxC is that while a single C file is nice and simple, for a large project it can compile pretty slowly, especially when optimizing. I’ve generated C files that take clang and gcc minutes to compile (others have seen even worse ). This is unfair to those compilers, of course, since C compilers aren’t expected to parallelize inside each source file. We could look into optionally emitting separate files in wasm2c to avoid that.

API questions

I don’t have many thoughts on this side as it isn’t really my area. I mostly just left the current wasm2c C API as it is, which is where those Z_ name manglings come from. You can use WasmBoxC right now, but you need to work at a fairly low level in C to bridge your code and the sandboxed code. Higher-level things are possible: TheRLBox people have done a lot of interesting work in this space on automation, tainting, etc., and the Sandboxed API is another approach that provides convenient and safe APIs for interacting with sandboxed code.

A specific example of something that could be improved is files. When doing sandboxing like this you don’t want to give the sandboxed code file access, and the current implementation of course prevents that. If you sandbox a library that does need file data, the usual pattern is to malloc inside the sandbox and copy data over to it (as is done in the gist example of zlib ). There may be better things we can do, like bundling of data files.

On the topic of files, you can enable support for them optionally, which is how the benchmarks were run: Lua needs to load files, and SQLite also needs to write and seek in its database file. Currently when emcc is used with the WASM2C flag it will fully sandbox things if you build a library (for which as we saw you should pass --no-entry ), while if you build a normal executable with a main() then it gives the code normal OS access. We probably want to think about a better design here, but meanwhile this has been useful for testing and benchmarking, and also as a way to get portable builds .

Safe languages

WasmBoxC may be especially interesting with a memory-safe language like Rust, Swift, or Go because it can help remove the unsafety of C libraries which would otherwise mar the security of an application written almost entirely in a safe language.

In fact, these languages generally have a bindings generator for C libraries. Perhaps the build system could have an option to WasmBox a library and generate appropriate bindings for it automatically? That would be as convenient as the hypothetical --sandbox flag we wished for in the introduction! If it’s that easy, and since the WasmBoxC overhead is so low, perhaps sandboxing of C libraries could become a common practice in safe languages? (I’d love to collaborate on this if someone is interested!)

Another interesting possibility here might be to compile not to C but to the memory-safe language itself, see for example wasm-to-rust . That would let you do everything in the safe language with no C at all. That wouldn’t make the sandboxing itself any more secure, but it could prevent bugs in the interface code, and might be more convenient.

Conclusion

The title of this post describes WasmBoxC as “ Simple, Easy, and Fast VM-less Sandboxing ”. Simplicity is the key from which all the rest of the benefits come: WasmBoxC compiles to a simple subset of C and it does so in a simple way, using existing wasm toolchains. By compiling to C we can run a C compiler on it which is easy and gives us fast code. It also makes it easy to communicate with the sandbox. And by using existing wasm toolchains we have well-tested and convenient compilers that let us build and sandbox code in an easy way.

I’m not aware of a sandboxing approach that is easier to use than WasmBoxC. And it’s interesting that we can reach just 14%-42% overhead with such a simple approach! It turns out that we can be both simple and fast.

Thank youto Sam Clegg, John Regehr, Ben Smith, and Patrick Walton for comments on drafts of this post, and a special thank you to Ben Smith for writing wasm2c and to Andre Weissflog for an inspiring tweet .