wasmtime / issue #14318 Preview1 adapter still imports re... · git-wasmtime

Stream: git-wasmtime

Topic: wasmtime / issue #14318 Preview1 adapter still imports re...

Wasmtime GitHub notifications bot (Sep 11 2026 at 18:00):

carsonfarmer opened issue #14318:

Test Case

A finite Go component calls an import returning a 64 KiB string 4,096 times, retains eight results, and asserts that GC ran. A long-lived time.AfterFunc initializes Go's timer poller. The reproducer contains no application storage or server dependencies.

Source: https://github.com/carsonfarmer/componentize-go/tree/cf/gc-clock-timer-regression/tests/fixtures/memory-pressure

Steps to Reproduce
git clone --branch cf/gc-clock-timer-regression https://github.com/carsonfarmer/componentize-go.git
cd componentize-go
GOTOOLCHAIN=go1.27.1 CARGO_BUILD_JOBS=2 cargo test -p tests fixture_memory_pressure -- --nocapture
The fixture runner builds the component using the repository's embedded preview1 adapter. The test fails with the unchanged embedded adapter. It also fails with the official Wasmtime 48.0.2 reactor adapter.

To distinguish the two remaining paths, the wall-clock-only adapter on carsonfarmer/wasmtime:cf/adapter-safe-clock fixes the realtime failure but still traps on timer polling. Adding the single immediate-timer case fixes that second failure. Both comparisons were reproduced with Go 1.26.3 and 1.27.1.

Expected Results

The allocation loop completes and verifies that garbage collection occurred, without importing host functions during cabi_realloc.

Actual Results

Two remaining paths can leave the component during canonical allocation:

clock_time_get → runtime.walltime1 → runtime.gcMarkTermination → cabi_realloc

With realtime caching fixed: poll_oneoff → runtime.netpoll → runtime.startTheWorldWithSema → runtime.gcStart → cabi_realloc

Both trap with cannot leave component instance.

Versions and Environment

macOS, arm64. Official Wasmtime 48.0.2 reactor adapter; componentize-go 0.4.2 / main 5b8eef1155ae8ea7f8bde9181cf72a568fd6eab3; Go 1.26.3 and 1.27.1. The fixture runner uses Wasmtime 48.0.1. The affected adapter code is also present on Wasmtime main 2607a7a31732313bce8aec80de9744221c62bc8f.

Extra Info

This follows #13563, which added monotonic-clock pausing for https://github.com/bytecodealliance/componentize-go/issues/56, together with https://github.com/bytecodealliance/go-pkg/pull/7. The fixture extends the test introduced in https://github.com/bytecodealliance/componentize-go/pull/60.

As explained in #12112, the component import restriction is correct. The proposed adapter change avoids those imports; it does not relax the restriction or support arbitrary I/O during allocation.

Wasmtime GitHub notifications bot (Sep 11 2026 at 18:01):

carsonfarmer edited issue #14318:

Test Case

A finite Go component calls an import returning a 64 KiB string 4,096 times, retains eight results, and asserts that GC ran. A long-lived time.AfterFunc initializes Go's timer poller. The reproducer contains no application storage or server dependencies.

Source: https://github.com/carsonfarmer/componentize-go/tree/cf/gc-clock-timer-regression/tests/fixtures/memory-pressure

Steps to Reproduce
git clone --branch cf/gc-clock-timer-regression https://github.com/carsonfarmer/componentize-go.git
cd componentize-go
GOTOOLCHAIN=go1.27.1 CARGO_BUILD_JOBS=2 cargo test -p tests fixture_memory_pressure -- --nocapture
The fixture runner builds the component using the repository's embedded preview1 adapter. The test fails with the unchanged embedded adapter. It also fails with the official Wasmtime 48.0.2 reactor adapter.

To distinguish the two remaining paths, the wall-clock-only adapter on carsonfarmer/wasmtime:cf/adapter-safe-clock fixes the realtime failure but still traps on timer polling. Adding the single immediate-timer case fixes that second failure. Both comparisons were reproduced with Go 1.26.3 and 1.27.1.

Expected Results

The allocation loop completes and verifies that garbage collection occurred, without importing host functions during cabi_realloc.

Actual Results

Two remaining paths can leave the component during canonical allocation:

clock_time_get → runtime.walltime1 → runtime.gcMarkTermination → cabi_realloc

With realtime caching fixed: poll_oneoff → runtime.netpoll → runtime.startTheWorldWithSema → runtime.gcStart → cabi_realloc

Both trap with cannot leave component instance.

Versions and Environment

macOS, arm64. Official Wasmtime 48.0.2 reactor adapter; componentize-go 0.4.2 / main 5b8eef1155ae8ea7f8bde9181cf72a568fd6eab3; Go 1.26.3 and 1.27.1. The fixture runner uses Wasmtime 48.0.1. The affected adapter code is also present on Wasmtime main 2607a7a31732313bce8aec80de9744221c62bc8f.

Extra Info

This follows #13563, which added monotonic-clock pausing for https://github.com/bytecodealliance/componentize-go/issues/56, together with https://github.com/bytecodealliance/go-pkg/pull/7. The fixture extends the test introduced in https://github.com/bytecodealliance/componentize-go/pull/60.

As explained in #12112, the component import restriction is correct. The proposed adapter change avoids those imports; it does not relax the restriction or support arbitrary I/O during allocation.

Proposed fix: https://github.com/bytecodealliance/wasmtime/pull/14319. Companion regression: https://github.com/bytecodealliance/componentize-go/pull/78.

Wasmtime GitHub notifications bot (Sep 11 2026 at 18:08):

carsonfarmer edited issue #14318:

Go components can still trap with cannot leave component instance when garbage collection runs during cabi_realloc.

The pause introduced in #13563 covers monotonic clock reads, but GC also reads wall time and polls timers when restarting the scheduler. Those operations still call host imports, which are forbidden during cabi_realloc.

Reproduced with the Wasmtime 48.0.2 adapter and Go 1.26.3/1.27.1. The regression test fails without the fix and passes with it.

Fix: #14319. Follow-up to https://github.com/bytecodealliance/componentize-go/issues/56.

Last updated: Sep 20 2026 at 18:08 UTC