fix(madsim): drop the future within the scope of TaskEnterGuard when a task is aborted

[BugenZhao]

Signed-off-by: Bugen Zhao [email protected]

[kwannoel]

Ohhhh I see. The continue triggered us to directly proceed to the end of the loop, and call the destructor of runnable implicitly. Then it calls the drop for the struct, which spawned the thread.

Seems acceptable to me to drop it because task was cancelled / node was killed.

[kwannoel]

LGTM once pass CI. @wangrunji0408 could you take a look too?

[kwannoel]

I considered opposite cases like:

    #[test]
    fn spawn_futures() {
        let runtime = madsim::runtime::Runtime::new();
        runtime.block_on(async move {
            let rt = Runtime::new().unwrap();
            let join_handle = rt.spawn(async move {
                Handle.spawn(async move {
                    madsim::time::sleep(Duration::from_secs(1)).await;
                    println!("hi");
                });
                Handle.spawn(async move {
                    madsim::time::sleep(Duration::from_secs(1)).await;
                    println!("hi2")
                });
            });
            madsim::time::sleep(Duration::from_secs(10)).await;
            join_handle.await.unwrap();
        });
        panic!("break");
    }

In such cases, we will always have task context, since their parent future will be executing with task context, and the spawning steps happen sequentially. The newly spawned futures will be handled by Spawner, which will schedule these tasks using its sender, and let the main event loop inside the executor execute these.

So I think we just need to handle the case of node kill / task abort as was done in this PR.

[kwannoel]

Many thanks @BugenZhao!

[BugenZhao]

Also relax this check so that spawn during drop by killing a node won't panic.

[kwannoel]

Let's also bump the version of madsim to v0.2.31 before merging.

[BugenZhao]

Given that this PR has been tested in risingwavelabs/risingwave#18852 and RisingWave might be the largest project to adopt madsim, I believe we can safely merge this PR now.