I was trying to do a relatively simple change (make linked get_default
) to bail if _start or the default export aren't found instead of returning a method that does nothing: https://github.com/bytecodealliance/wasmtime/blob/d8d4bf81b260ea18f56475fbde96b6662a365cf8/crates/wasmtime/src/linker.rs#L1062
I was basically trying to finish up this long forgotten PR i had
Changing that last line in linker::get_default
with a bail like in that PR makes some reactor tests fail. E.g. this one:
#[test]
fn minimal_reactor() -> Result<()> {
let wasm = build_wasm("tests/wasm/minimal-reactor.wat")?;
let stdout = run_wasmtime(&[wasm.path().to_str().unwrap(), "--disable-cache"])?;
assert_eq!(stdout, "");
Ok(())
}
Adding a _start
function to minimal-reactor.wat
gets a complaint that it can't be both a command and a reactor.
(module
(func (export "_initialize"))
(func (export "_start"))
)
Result:
Caused by:
0: failed to instantiate "/var/folders/7y/bhkyc0053yb8cbd2swfz956wzsf2bs/T/.tmppNMSQK"
1: Program cannot be both a Command and a Reactor
So it turns out that's not the right way to patch the .wat
file, but it somehow expects some export in there.
This passes the test but looks suspicious (e.g. I wouldn't know how to generate this from a compiled program: e.g. rust), without manually editing the .wat file:
(module
(func (export "_initialize"))
(func (export ""))
)
Any advice on where I can read about reactors and how they are supposed to work?
(deleted)
But if I add a _start
function to the minimal reactor, the test fails as it says it can't be both a command and a reactor
not sure what the fix would be.
I'll see if other maintainers have an opinion about this.
I wanted to fix the PR above but it's unclear to me what correct is :grinning:.
Right now, get_default
is slightly confusing as it returns a newly created method if no suitable function is found when asking for a function from the module (probably not the caller's intent).
At the same time, in the reactor system there are some assumptions that the linker works like that in wasmtime. (and that _start
should not be present in reactors).
That said, the only slightly more serious issue I see is if someone brings their reactors to WASMTime from another VM ecosystem, they may not work, but it's really unclear what correct is here :grinning:
@Dan Gohman might know best how to fix this, but I suspect that the minimal reactor wants to keep working and it'd either be handled differently by the CLI or the get_default
function would continue doing what it does for reactors or something like that
What documentation there is for reactors is here: https://github.com/WebAssembly/WASI/blob/main/design/application-abi.md#current-unstable-abi
It hasn't been well developed yet, because in trying to figure out how reactors should work, we realized that what we really needed was for module linking etc. to subsume them.
Module linking is on its way to doing that now, but not all the pieces are in place yet. So reactors are what we have for now.
If we want get_default
to fail, I think what we'll need to do is make the wasmtime CLI gracefully handle get_default
failing -- that minimal_reactor
test case works by invoking the wasmtime CLI.
It may just be a matter of changing the get_default
call in src/commands/run.rs to do handle the failure case
Last updated: Nov 22 2024 at 16:03 UTC