Open nikomatsakis opened 4 months ago
This issue is intended for status updates only.
For general questions or comments, please contact the owner(s) directly.
Have been looking into different sandbox runtime choices. Here is a simple version of the comparison of three potential choices:
CAP_PERFMON
CAP_NET_ADMIN
CAP_SYS_ADMIN
or CAP_BPF
for loading eBPF programs, unless unprivileged eBPF is enabledmounts
in config.json
cargo run ubuntu:latest
to run docker image!?), and is not really portable.wasmtime --dir
.There are prior research on cross-over between each of these options. I've been busy these two weeks. Will update a more detail post for prior arts afterward.
The biggest challenge I am seeing now is spawning external processes. Most build script usages invoke some external binaries, like pkg-config
for building *-sys
crates, or protoc
for generating protobuf bindings. If process spawning is that common, we need to find a way to provide a fine-grained permission granting scheme. I don't want it to see an “all-or-nothing” mechanism when process spawning is needed.
The other huge headache is setting library search paths. We cannot know every possible path of system libraries ahead of time, but we need to grant access to the runtime.
None.
None.
Having a family urgency. I will be back after RustConf.
Building a workable version of wasm-based build script (not yet done). There are some technical difficulties. Not blocking but need to be addressed.
wasm32-wasip1
is not shipped by default; we need to rustup target add wasm32-wasip1
to build wasm.--target
flag is provided. Not providing --target wasm32-wasip1
means no target information available. AFAIK, this is already an issue of artifact-dependencies, e.g., https://github.com/rust-lang/cargo/issues/10444 and https://github.com/rust-lang/cargo/issues/11260.In order to make the development independent of Cargo, RUSTC_WRAPPER
was the first approach I tried, though it still lacks some extension point in Cargo:
RUSTC_WRAPPER
can only intercept rustc
invocations.target.runner
is available but for cargo test
and cargo run
only.Some compatibility issues came up when integrating with Cargo:
artifact-dependencies
can compile build script to any other platform, making it challenging to handle.None.
None.
Have a working-in-process pull request in https://github.com/weihanglo/cargo/pull/66.
Let me copy some texts from there :)
As you can see, we can easily swap to any sandbox runner with a custom target.
We use wasm32-wasip1
above as an example,
as it is the one with smaller footprint, very cross-platform, and pretty popular in the Rust community.
However, it turns out that `wasm32-wasip1 doesn't support POSIX process spawning.
Use cases of process spawning in build scripts are essential, such as
pkg-config
to find system libraries.cargo metadata
to get crate metadata.git
.According to the design axioms,
Restrcting process spawning is the top one axiom.
We made that with wasm32-wasip1
,
but have no way to opt-out.
This contradicts to the "ensuring -sys
crates be built" axiom.
As a result, it is unlikely to use wasm32-wasip1
as a default sandbox environment with this experiment.
Have talked to some other folks, there are some potential route we could take if we chose wasi as a default sandbox environment.
build-rs
crate to the rescueThe Cargo team recently adopted the build-rs
crate.
It is going to be the official crate providing API for writing build scripts.
We could take the advantage of it, telling everyone instead of using std::process::Command
, use something like build_rs::Command
. So that Cargo could have a full control over how a build script spawning processes.
While it sounds ideal, this doesn't help the current situation because
build-rs
in build scripts may increase build time; people may refuse to use.There was a discussion in the GSoC "sandboxed proc-macro with wasm" about shipping a custom verion of the standard library for sandboxed wasm. For Cargo's build script, we could potentially ship a wasm32-wasi-cargo
target. The std in this custom target could intercept any exec
call or process spawning. Then it calls back to the host process (which is Cargo in our case) to determine how to handle process execution. The host process could either reject, or run the external program and post back the result.
This idea sounds pretty hacky and need more investigations of the communication mechanism between the host process and the wasm runtime. Perhaps via sockets, The WebAssembly Component Model, other host function call mechanism. There is also WASIX project which supprots fork
/exec
though it is currently not a WASI standard not even a proposal.
Since one of major design space is the user interafce of sandbox configuration, we could leave off sandbox runtimes and explore more on the configuration side.
We could, for example, use docker or eBPF as a temporary default runtime, and explore how the configuration should look like. We may want to take a look at the configuration of Cackle-rs as a starting point. By doing so, we wouldn't block on waiting for wasm runtime to being more mature.
Unfortunately, no significant progress has been made. However, we received feedback on the previous experiment: https://github.com/weihanglo/cargo/pull/66.
Alex, who has been working on WebAssembly and is also an honorable Rust project contributor, expressed interest in adding POSIX process support to WASI. However, according to their comments, this seems unlikely in the near future. While the experiment with a Cargo-flavored WASI standard library remains incomplete, it is still on my to-do list. At the very least, I want to give it a try and see how gnarly it would be.
In this comment, Ralf pointed out that supporting process spawning may be somewhat pointless, as one of WASI's key selling points is avoiding the execution of external processes altogether. To me, this feels more like a UX design issue. For example, VS Code prompts users to "trust" a project when it is opened, although most users simply click "trust." With better UX design—such as interactive shell sessions for crate approvals or pre-approval lists from a company (similar to how cargo-vet operates)—an opt-in mechanism would be a better default than allowing everything by default.
Samuel shared a project called build-wrap, which leverages Bubblewrap. While not cross-platform, the configuration interface of Bubblewrap is inspiring and worth exploring, particularly for how it sandboxes system libraries and binaries.
Jeff suggested using the WebAssembly Component Model to allow interaction with a set of well-known programs (e.g., pkg-config
and cc
), with the host exposing an interface via Wasm components to build scripts. This would enable build scripts to be written in plain Rust while aligning with the use cases of the Wasm Component Model. This idea is particularly promising after the integration of build-rs into Cargo as a workspace member. While build-rs is a lower-level crate, it opens doors to integrating future build script helpers or interfaces.
Despite the valuable feedback and interest, after discussions with the Cargo team, this initiative will not continue as a project goal for 2025H1. Instead, there are alternative approaches worth exploring that may require less effort or depend less on parallel developments in other technologies:
Invest in system-deps
: This crate provides a declarative way to configure system dependencies. Building C (system) libraries is a key use case for build scripts and one of the reasons they exist. If system-deps
became a native Cargo solution, many packages could transition to being build-script-free. This issue explores how *-sys
crates are used and what is needed to increase adoption.
Cackle-like approaches: Cackle is a code checker that analyzes API usage across dependencies by wrapping rustc
and the linker to perform API analysis on object and debug info files. For build scripts, APIs like std::net
or std::process
could be disabled by default. While this is not a true sandboxing mechanism, Cackle optionally provides Bubblewrap as a sandbox environment, similar to the earlier-mentioned build-wrap.
A fully sandboxed build environment: Instead of partial sandboxing, we could explore integrating a fully sandboxed build environment, akin to how the Haskell tool Stack operates. Stack natively supports Docker with commands like stack docker
and also integrates with the Nix build system. However, this approach has significant compatibility challenges, such as limited Windows support and the added complexity of learning Docker or similar tools. While Haskell's Cabal system offers sandboxing or Nix-style local builds, it is more analogous to Python's virtualenv
than the sandboxing we're discussing here.
We want to explore various possibilities beyond the Wasm sandbox environment. That said, as mentioned earlier, personally I am still interested in experimenting further with Wasm. Sorry for not making more progress for the past month.
Summary
Explore different strategies for sandboxing build script executions in Cargo.
Tasks and status