"hello world" example with the wasi crate doesn't work

[yoshuawuyts]

Repro

If we run the following commands to build the wasi crate:

cargo install cargo-component
git clone https://github.com/bytecodealliance/wasi
cd wasi
cargo component build --example hello-world

This results in the following output:

yosh@MacBook-Pro wasi % cargo component run --example hello-world
   Compiling wit-bindgen v0.19.1
   Compiling bitflags v2.4.2
   Compiling wasi v0.12.1+wasi-0.2.0 (/Users/yosh/Code/wasi)
    Finished dev [unoptimized + debuginfo] target(s) in 1.76s
Error: failed to run main module `/Users/yosh/Code/wasi/target/wasm32-wasi/debug/examples/hello-world.wasm`

Caused by:
    0: failed to instantiate "/Users/yosh/Code/wasi/target/wasm32-wasi/debug/examples/hello-world.wasm"
    1: unknown import: `wasi:io/[email protected]::[resource-drop]output-stream` has not been defined

Manual build instructions

The wasi crate suggests the following instructions to get it to build correctly:

# build the example for WASI 0.1
cargo build --target wasm32-wasi --example hello-world

# fetch the WASI 0.1 -> WASI 0.2 adapter
curl -LO https://github.com/bytecodealliance/wasmtime/releases/download/v17.0.0/wasi_snapshot_preview1.command.wasm

# convert the example into a WASI 0.2 component
cargo install wasm-tools
wasm-tools component new target/wasm32-wasi/example/hello-world.wasm \
    --adapt ./wasi_snapshot_preview1.command.wasm \
    -o hello-world.wasm

# run the component
wasmtime run hello-world.wasm

This results in the following output:

Hello, world!

Expected behavior

From the outset I was expecting this to just work? Is it that cargo run --example isn't implemented yet? It does seem to build the wasm32-wasi example correctly. Or is something else perhaps not working when it should?

[peterhuene]

To prevent cargo-component from componentizing any ol' wasm32 target it encounters, it expects Cargo.toml to have the [package.metadata.component] section to treat it as a component rather than a module. You'll notice it is missing the Creating component message in the cargo component build output.

If you add [package.metadata.component] to Cargo.toml, you'll get errors when parsing the WIT, as cargo-component does not do WIT dependency discovery via a deps directory, instead expects entries in [package.metadata.component.target.dependencies].

So Cargo.toml would need:

[package.metadata.component.target.dependencies]
"wasi:http" = { path = "wit/deps/http" }
"wasi:random" = { path = "wit/deps/random" }
"wasi:cli" = { path = "wit/deps/cli" }
"wasi:io" = { path = "wit/deps/io" }
"wasi:clocks" = { path = "wit/deps/clocks" }
"wasi:filesystem" = { path = "wit/deps/filesystem" }
"wasi:sockets" = { path = "wit/deps/sockets" }

However, there appears to be breaking changes in the latest wit-bindgen with regards to runtime types, so you'd need to drop the version down to 0.18.0 to be compatible with the latest cargo-component (I should do a cargo-component release to bump wit-bindgen to latest).

With the above in place, you should now see the following for cargo component run --example hello-world:

Error: failed to run main module `/Users/peterhuene/tmp/wasi/target/wasm32-wasi/debug/examples/hello-world.wasm`

Caused by:
    exported instance `wasi:cli/[email protected]` not present

This, finally, is a bug in cargo-component, where it is not treating an example as a bin, and thus is using the built-in reactor adapter rather than the command adapter.

It's possible to workaround that with:

[package.metadata.component]
adapter = "wasi_snapshot_preview1.command.wasm"

But it should work out of the box.

[alexcrichton]

it expects Cargo.toml to have the [package.metadata.component] section to treat it as a component rather than a module

In this case with no WIT internally as it's all coming from the wasi crate itself could this logic change to: if no [package.metadata.component] is specified then parse the output wasm module and if any component-type* custom section is found then do the componentization process as-if an empty section was specified?

However, there appears to be breaking changes in the latest wit-bindgen with regards to runtime types

Oh dear, do you mean breakage from 0.18-to-0.19 or something that needs a yank/republish on crates.io?

(also I'm happy to cc you on more changes to wit-bindgen if you'd like for possible breakage with cargo-component)

[peterhuene]

In this case with no WIT internally as it's all coming from the wasi crate itself could this logic change to: if no [package.metadata.component] is specified then parse the output wasm module and if any component-type* custom section is found then do the componentization process as-if an empty section was specified?

I agree that componetizing a module based on the presence of the component-type custom section rather than the presence of [package.metadata.component] makes more sense; I can make that change.

Right now, if there are no settings under [package.metadata.component.target], cargo-component will probe for a wit directory to generate bindings automatically; hence why it's generating bindings for the bindings world in wit/wasi-crate.wit here.

I'm not entirely sure I would want to change that behavior, as most of the time the presence of the wit directory indicates a local target for the component being built.

Oh dear, do you mean breakage from 0.18-to-0.19 or something that needs a yank/republish on crates.io?

A breakage from 0.18 to 0.19; it's just cargo-component that's out of date, as it uses 0.18.0 to generate bindings ahead-of-time, but Cargo.toml might reference 0.19.1 solely for the runtime types defined in wit-bindgen.

An example breakage from 0.18.0 to 0.19.1:

error[E0599]: no function or associated item named `into_handle` found for struct `Resource<_>` in the current scope
     --> src/bindings.rs:14337:40
      |
14337 |             wit_bindgen::rt::Resource::into_handle(self.handle)
      |                                        ^^^^^^^^^^^
      |                                        |
      |                                        function or associated item not found in `Resource<_>`
      |                                        help: there is an associated function with a similar name: `handle`

Ideally it would be nice if we had a mode where we could generate the runtime types (and cabi_* exports) as part of bindings generation so that the user's Cargo.toml doesn't depend on wit-bindgen at all.

[alexcrichton]

Ah I see good point!

One suggestion would be to update cargo component to generate wit-bindgen = "0.18.0" in the manifest, or whatever matches the built-in generator of cargo component. That being said that won't play well with upgrades over time, for example taking a very new cargo component and building an older project.

One fix for that is to then do what you mention, generate Reource<T> and bits into the file itself. That has problems though we just discovered at least with cabi_realloc where we want multiple wit-bindgen-using crates to link together and they can't all define cabi_realloc so we need to use weak symbols, and then that's not stable in Rust.

I think the best solution might be to define "perma stable" APIs, like how the runtime links cabi_realloc and the run_ctors_once thing. Less stable things, like Resource<T> should go into the generated bindings.

So how about this:

• As a near-term fix, cargo component updates to 0.19.x for wit-bindgen generation
• As a long-term fix, wit-bindgen generates the Resource<T> wrapper and any "less stable" bits into the generated bindings
• As a long-term fix, wit-bindgen considers its API surface are "perma stable" as future versions of wit-bindgen runtime should work with old versions of wit-bindgen generators

[peterhuene]

Your plan sounds great to me!