-
Notifications
You must be signed in to change notification settings - Fork 18
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
wasmer panic when await multiple async Host function with join #194
Comments
Thanks for the detailed report! I've tried to create a test case to reproduce it myself, but so far it's not failing for me. See: #195 Do you maybe see if I missed some detail from your report? |
Oh, and maybe it matters which |
Hi @arendjr , thanks a lot for your quick response! After some further investigation, I finally find out the core difference between your test case in #195 and my code:
The code for mod wasi_spec;
use anyhow::Result;
use crate::wasi_spec::bindings::Runtime;
use std::sync::Mutex;
pub static GLOBAL_STATE: Mutex<u32> = Mutex::new(0);
const WASM_BYTES: &'static [u8] =
include_bytes!("../../example-plugin/target/wasm32-wasi/debug/example_plugin.wasm");
fn new_runtime() -> Result<Runtime> {
let rt = Runtime::new(WASM_BYTES)?;
rt.init()?;
Ok(rt)
}
#[tokio::main]
async fn main() -> Result<()> {
println!("Hello, world!");
let rt = new_runtime()?;
let response = rt.make_parallel_requests().await?;
// assert_eq!(response, r#"{"status":"confirmed"}"#.to_string());
assert_eq!(response, response);
Ok(())
} So I guess the issue happens with some Tokio related init/uninit steps? Woud you please help to verify it? |
I think it has something to do with multithreading: By setting
I can reproduce this panic with test If I set main to
It also works as expected. Hopefully that could narrow down the issue a little bit |
Thank you so much! This points into a direction I was already having suspicions about. I have not yet fully confirmed this, but I highly suspect the issue is that callbacks from Tokio can come from any thread, whereas the code running inside WebAssembly assumes a single-threaded context. That assumption is violated if you call back into the sandbox from multiple threads simultaneously, which may happen if parallel requests are triggered inside the sandbox and the callbacks to those can come back from multiple threads at once. At least, that would explain why Tokio's "current_thread" flavor has no problems and why I hadn't run into this before without parallel requests. I had always made the assumption that Wasmer would be responsible for synchronizing callbacks back into the sandbox, but apparently it doesn't do this. I suspect the migration to Wasmer 3 might resolve this, since it changes the At least we have a work-around now, and I have some more pointers to attempt a real fix. Thanks again! |
Thanks a lot for your quick response again. I'm glad the info I gave could provide some help. Currently I can play with the work-around and continue my development, looking forward to hear your good news soon! |
Hi @arendjr I tried to fix this issue with a local modification fp-bindgen on my own, and I did upgrade to Wamser 4 (based on the work Roy Jacobs roy.jacobs@gmail.com did on wasmer3 branch), with a quick and dirty hack, I made it work just like what it is for wasmer2:
One issue with Wasmer4 is But that's not my key point here, my question is, even if we make it work with Wamser4, it seems like we still have to async function panic issue, what further action should I take, try to improve Wasmer4 related changes, or try another direction? By saying 'another direction', I have a wild guess: by checking the file |
|
I may need to spend some more time to look deeper into this issue, but from my cursory understanding, when you say:
I would suspect the use of What I think is the right approach here, is for us to implement some mechanism to run the WASM environment in a single thread (may be the one from which it is created, or a dedicated one) and use a channel to make sure callbacks are all processed by that same thread. I haven’t yet looked at |
hi @arendjr I followed your advice and dis some resarch, and come up with a primitive fix: it does have performance issue, but at least I could get correct results in multi-thread tokio runtime without any panic. Would you please help to review this change? And feel free to copy it partially if you some pieces are useful but unable to merge it due to whatever reason. If you believe that's the right direction, I could spend more efforts to improve it. the PR: #210 |
Sorry, I was on holiday. I no longer work at Fiberplane, so I can't really help with this anymore. |
Background
I'd like to build a service with Wasm and fp-bindgen, with both import and export function as async functions. In order to make guest functions run concurrently, I'd like to involve some function like
tokio::join
orfutures::future::join
to run Host's async function concurrently in guest, but run with a panic consistantly (80%)Here is the panic message:
The related source code
Quetion
Is it possible to run host's async functions concurrently within guest's async function with
join
?Code details
My fp-bindgen protocol:
Guest code:
Host code:
implementation of exported functions: just two empty async functions
main function:
The text was updated successfully, but these errors were encountered: