/runwasi

Facilitates running Wasm / WASI workloads managed by containerd

Primary LanguageRustApache License 2.0Apache-2.0

runwasi logo light mode runwasi logo dark mode

runwasi

Warning: Alpha quality software, do not use in production.

This is a project to facilitate running wasm workloads managed by containerd either directly (ie. through ctr) or as directed by Kubelet via the CRI plugin. It is intended to be a (rust) library that you can take and integrate with your wasm host. Included in the repository is a PoC for running a plain wasi host (ie. no extra host functions except to support wasi system calls).

Community

Usage

runwasi is intended to be consumed as a library to be linked to from your own wasm host implementation.

There are two modes of operation supported:

  1. "Normal" mode where there is 1 shim process per container or k8s pod.
  2. "Shared" mode where there is a single manager service running all shims in process.

In either case you need to implement a trait to teach runwasi how to use your wasm host.

There are two ways to do this:

  • implementing the sandbox::Instance trait
  • or implementing the container::Engine trait

The most flexible but complex is the sandbox::Instance trait:

pub trait Instance {
    /// The WASI engine type
    type Engine: Send + Sync + Clone;

    /// Create a new instance
    fn new(id: String, cfg: Option<&InstanceConfig<Self::E>>) -> Self;
    /// Start the instance
    /// The returned value should be a unique ID (such as a PID) for the instance.
    /// Nothing internally should be using this ID, but it is returned to containerd where a user may want to use it.
    fn start(&self) -> Result<u32, Error>;
    /// Send a signal to the instance
    fn kill(&self, signal: u32) -> Result<(), Error>;
    /// Delete any reference to the instance
    /// This is called after the instance has exited.
    fn delete(&self) -> Result<(), Error>;
    /// Wait for the instance to exit
    /// The waiter is used to send the exit code and time back to the caller
    /// Ideally this would just be a blocking call with a normal result, however
    /// because of how this is called from a thread it causes issues with lifetimes of the trait implementer.
    fn wait(&self, waiter: &Wait) -> Result<(), Error>;
}

The container::Engine trait provides a simplified API:

pub trait Engine: Clone + Send + Sync + 'static {
    /// The name to use for this engine
    fn name() -> &'static str;
    /// Run a WebAssembly container
    fn run_wasi(&self, ctx: &impl RuntimeContext, stdio: Stdio) -> Result<i32>;
    /// Check that the runtime can run the container.
    /// This checks runs after the container creation and before the container starts.
    /// By it checks that the wasi_entrypoint is either:
    /// * a file with the `wasm` filetype header
    /// * a parsable `wat` file.
    fn can_handle(&self, ctx: &impl RuntimeContext) -> Result<()> { /* default implementation*/ }
}

After implementing container::Engine you can use container::Instance<impl container::Engine>, which implements the sandbox::Instance trait.

To use your implementation in "normal" mode, you'll need to create a binary which has a main that looks something like this:

use containerd_shim as shim;
use containerd_shim_wasm::sandbox::{ShimCli, Instance}

struct MyInstance {
    // ...
}

impl Instance for MyInstance {
    // ...
}

fn main() {
    shim::run::<ShimCli<MyInstance>>("io.containerd.myshim.v1", opts);
}

or when using the container::Engine trait, like this:

use containerd_shim as shim;
use containerd_shim_wasm::{sandbox::ShimCli, container::{Instance, Engine}}

struct MyEngine {
    // ...
}

impl Engine for MyEngine {
    // ...
}

fn main() {
    shim::run::<ShimCli<Instance<Engine>>>("io.containerd.myshim.v1", opts);
}

Note you can implement your own ShimCli if you like and customize your wasm engine and other things. I encourage you to checkout how that is implemented.

The shim binary just needs to be installed into $PATH (as seen by the containerd process) with a binary name like containerd-shim-myshim-v1.

For the shared mode:

use containerd_shim_wasm::sandbox::{Local, ManagerService, Instance};
use containerd_shim_wasm::services::sandbox_ttrpc::{create_manager, Manager};
use std::sync::Arc;
use ttrpc::{self, Server};
/// ...

struct MyInstance {
    /// ...
}

impl Instance for MyInstance {
    // ...
}

fn main() {
    let s: ManagerService<Local<MyInstance>> =
        ManagerService::new(Engine::new(Config::new().interruptable(true)).unwrap());
    let s = Arc::new(Box::new(s) as Box<dyn Manager + Send + Sync>);
    let service = create_manager(s);

    let mut server = Server::new()
        .bind("unix:///run/io.containerd.myshim.v1/manager.sock")
        .unwrap()
        .register_service(service);

    server.start().unwrap();
    let (_tx, rx) = std::sync::mpsc::channel::<()>();
    rx.recv().unwrap();
}

This will be the host daemon that you startup and manage on your own. You can use the provided containerd-shim-myshim-v1 binary as the shim to specify in containerd.

Shared mode requires precise control over real threads and as such should not be used with an async runtime.

Check out these projects that build on top of runwasi:

Components

  • containerd-shim-[ wasmedge | wasmtime | wasmer ]-v1

This is a containerd shim which runs wasm workloads in WasmEdge or Wasmtime or Wasmer. You can use it with containerd's ctr by specifying --runtime=io.containerd.[ wasmedge | wasmtime | wasmer ].v1 when creating the container. And make sure the shim binary must be in $PATH (that is the $PATH that containerd sees). Usually you just run make install after make build.

build shim with wasmedge we need install library first

This shim runs one per pod.

  • containerd-shim-[ wasmedge | wasmtime | wasmer ]d-v1

A cli used to connect containerd to the containerd-[ wasmedge | wasmtime | wasmer ]d sandbox daemon. When containerd requests for a container to be created, it fires up this shim binary which will connect to the containerd-[ wasmedge | wasmtime | wasmer ]d service running on the host. The service will return a path to a unix socket which this shim binary will write back to containerd which containerd will use to connect to for shim requests. This binary does not serve requests, it is only responsible for sending requests to the containerd-[ wasmedge | wasmtime | wasmer ]d daemon to create or destroy sandboxes.

  • containerd-[ wasmedge | wasmtime | wasmer ]d

This is a sandbox manager that enables running 1 wasm host for the entire node instead of one per pod (or container). When a container is created, a request is sent to this service to create a sandbox. The "sandbox" is a containerd task service that runs in a new thread on its own unix socket, which we return back to containerd to connect to.

The Wasmedge / Wasmtime / Wasmer engine is shared between all sandboxes in the service.

To use this shim, specify io.containerd.[ wasmedge | wasmtime | wasmer ]d.v1 as the runtime to use. You will need to make sure the containerd-[ wasmedge | wasmtime | wasmer ]d daemon has already been started.

Contributing

To begin contributing, learn to build and test the project or to add a new shim please read our CONTRIBUTING.md

Demo

Installing the shims for use with Containerd

Make sure you have installed dependencies and install the shims:

make build
sudo make install

Note: make build will only build one binary. The make install command copies the binary to $PATH and uses symlinks to create all the component described above.

Build the test image and load it into containerd:

make test-image
make load

Demo 1 using container image that contains a Wasm module.

Run it with sudo ctr run --rm --runtime=io.containerd.[ wasmedge | wasmtime | wasmer ].v1 ghcr.io/containerd/runwasi/wasi-demo-app:latest testwasm /wasi-demo-app.wasm echo 'hello'. You should see some output repeated like:

sudo ctr run --rm --runtime=io.containerd.wasmtime.v1 ghcr.io/containerd/runwasi/wasi-demo-app:latest testwasm

This is a song that never ends.
Yes, it goes on and on my friends.
Some people started singing it not knowing what it was,
So they'll continue singing it forever just because...

This is a song that never ends.
Yes, it goes on and on my friends.
Some people started singing it not knowing what it was,
So they'll continue singing it forever just because...

(...)

To kill the process, you can run in other session: sudo ctr task kill -s SIGKILL testwasm.

The test binary supports commands for different type of functionality, check crates/wasi-demo-app/src/main.rs to try it out.

Demo 2 using OCI Images with custom WASM layers

The previous demos run with an OCI Container image containing the wasm module in the file system. Another option is to provide a cross-platform OCI Image that that will not have the wasm module or components in the file system of the container that wraps the wasmtime/wasmedge process. This OCI Image with custom WASM layers can be run across any platform and provides for de-duplication in the Containerd content store among other benefits. To build OCI images using your own images you can use the oci-tar-builder

To learn more about this approach checkout the design document.

Note: This requires containerd 1.7.7+ and 1.6.25+. If you do not have these patches for both containerd and ctr you will end up with an error message such as mismatched image rootfs and manifest layers at the import and run steps. Latest versions of k3s and kind have the necessary containerd versions.

Build and import the OCI image with WASM layers image:

make test-image/oci
make load/oci

Run the image with sudo ctr run --rm --runtime=io.containerd.[ wasmedge | wasmtime | wasmer ].v1 ghcr.io/containerd/runwasi/wasi-demo-oci:latest testwasmoci

sudo ctr run --rm --runtime=io.containerd.wasmtime.v1 ghcr.io/containerd/runwasi/wasi-demo-oci:latest testwasmoci wasi-demo-oci.wasm echo 'hello'
hello
exiting