Stream: git-wasmtime

Topic: wasmtime / issue #12128 component-model-async: Cannot ree...

Wasmtime GitHub notifications bot (Dec 05 2025 at 16:05):

alexcrichton opened issue #12128:

This test case:

(component
  (component $a
    (core module $a
      (import "" "yield" (func $yield (result i32)))

      (func (export "yield-loop") (result i32)
        ;; simulate `waitable-set.poll` with a yield loop
        (loop
          call $yield
          drop
          br 0
        )
        unreachable
      )

      ;; not reached
      (func (export "callback") (param i32 i32 i32) (result i32) unreachable)

      (func (export "noop"))
    )
    (core func $yield (canon thread.yield))
    (core instance $a (instantiate $a
      (with "" (instance
        (export "yield" (func $yield))
      ))
    ))
    (func (export "yield-loop")
      (canon lift
        (core func $a "yield-loop")
        async
        (callback (func $a "callback"))
      )
    )
    (func (export "noop") (canon lift (core func $a "noop")))
  )
  (instance $a (instantiate $a))

  (component $b
    (import "yield-loop" (func $yield-loop))
    (import "noop" (func $noop))

    (core func $yield-loop (canon lower (func $yield-loop) async))
    (core func $noop (canon lower (func $noop)))

    (core module $b
      (import "" "yield-loop" (func $yield-loop (result i32)))
      (import "" "noop" (func $noop))

      (func (export "run")
        ;; call `yield-loop`, double-check it's in the "started" state.
        call $yield-loop
        i32.const 0xf
        i32.and
        i32.const 1
        i32.ne
        if unreachable end

        ;; now try to reenter the other component with some other function.
        call $noop
      )
    )
    (core instance $b (instantiate $b
      (with "" (instance
        (export "yield-loop" (func $yield-loop))
        (export "noop" (func $noop))
      ))
    ))
    (func (export "run") (canon lift (core func $b "run")))
  )
  (instance $b (instantiate $b
    (with "yield-loop" (func $a "yield-loop"))
    (with "noop" (func $a "noop"))
  ))
  (export "run" (func $b "run"))
)

(assert_return (invoke "run"))

fails with:

$ WASMTIME_BACKTRACE_DETAILS=1 wasmtime wast foo.wast -W component-model-async
Error: failed to run script file 'foo.wast'

Caused by:
    0: failed directive on foo.wast:77
    1: error while executing at wasm backtrace:
           0:    0x185 - <unknown>!<wasm function 8>
           1:    0x333 - wasm-function[2]
                           at ./foo.wast:59:9
    2: wasm trap: wasm `unreachable` instruction executed

Specifically $yield-loop is, well, yielding, and then the second call to $noop I believe is dying in the fused adapter. My best guess is due to component reentrance, but I'm not 100% certain since errors from the fused adapter are pretty opaque.

Is this intended behavior? Accidental behavior?

Wasmtime GitHub notifications bot (Dec 05 2025 at 16:05):

alexcrichton added the wasm-proposal:component-model-async label to Issue #12128.

Wasmtime GitHub notifications bot (Dec 05 2025 at 21:02):

alexcrichton edited issue #12128:

This test case:

(component
  (component $a
    (core module $a
      (import "" "yield" (func $yield (result i32)))

      (func (export "yield-loop") (result i32)
        ;; simulate `waitable-set.poll` with a yield loop
        (loop
          call $yield
          drop
          br 0
        )
        unreachable
      )

      ;; not reached
      (func (export "callback") (param i32 i32 i32) (result i32) unreachable)

      (func (export "noop"))
    )
    (core func $yield (canon thread.yield))
    (core instance $a (instantiate $a
      (with "" (instance
        (export "yield" (func $yield))
      ))
    ))
    (func (export "yield-loop") async
      (canon lift
        (core func $a "yield-loop")
        async
        (callback (func $a "callback"))
      )
    )
    (func (export "noop") (canon lift (core func $a "noop")))
  )
  (instance $a (instantiate $a))

  (component $b
    (import "yield-loop" (func $yield-loop async))
    (import "noop" (func $noop))

    (core func $yield-loop (canon lower (func $yield-loop) async))
    (core func $noop (canon lower (func $noop)))

    (core module $b
      (import "" "yield-loop" (func $yield-loop (result i32)))
      (import "" "noop" (func $noop))

      (func (export "run")
        ;; call `yield-loop`, double-check it's in the "started" state.
        call $yield-loop
        i32.const 0xf
        i32.and
        i32.const 1
        i32.ne
        if unreachable end

        ;; now try to reenter the other component with some other function.
        call $noop
      )
    )
    (core instance $b (instantiate $b
      (with "" (instance
        (export "yield-loop" (func $yield-loop))
        (export "noop" (func $noop))
      ))
    ))
    (func (export "run") async (canon lift (core func $b "run")))
  )
  (instance $b (instantiate $b
    (with "yield-loop" (func $a "yield-loop"))
    (with "noop" (func $a "noop"))
  ))
  (export "run" (func $b "run"))
)

(assert_return (invoke "run"))

fails with:

$ WASMTIME_BACKTRACE_DETAILS=1 wasmtime wast foo.wast -W component-model-async
Error: failed to run script file 'foo.wast'

Caused by:
    0: failed directive on foo.wast:77
    1: error while executing at wasm backtrace:
           0:    0x185 - <unknown>!<wasm function 8>
           1:    0x333 - wasm-function[2]
                           at ./foo.wast:59:9
    2: wasm trap: wasm `unreachable` instruction executed

Specifically $yield-loop is, well, yielding, and then the second call to $noop I believe is dying in the fused adapter. My best guess is due to component reentrance, but I'm not 100% certain since errors from the fused adapter are pretty opaque.

Is this intended behavior? Accidental behavior?

Wasmtime GitHub notifications bot (Dec 05 2025 at 21:03):

alexcrichton commented on issue #12128:

Ok in discussion with @lukewagner the conclusion is:

• Prior to #12043 this test should block forever. Notably the noop call sould be blocked on backpressure being turned on as part of yield-loop
• Post-#12043 this test should pass (with some minor async edits I've applied to the OP now)

With #12043 as-is this test still traps in pretty much the same place, so this is something we'll want to fix.

Wasmtime GitHub notifications bot (Dec 05 2025 at 21:03):

alexcrichton assigned dicej to issue #12128.

Last updated: Dec 06 2025 at 07:03 UTC