Files
base/plugins/gojavm/module.go
zeekayandGitHub 65ef705610 feat(base): pluggable extension runtimes — wazero + v8go + goja + native + bench (#3)
* feat(base): extension runtime interface + native impl

Foundation for pluggable extension runtimes. One Runtime interface
that wazero, v8go, goja, and native-Go all satisfy. Extensions
declare their preferred runtime in extension.json; the loader
dispatches.

This sits alongside plugins/jsvm — goja-via-the-existing-jsvm-path
remains the default for .base.js hooks. The new path is for users
who want hard-sandboxed WASM extensions or full V8 perf.

Files:
- plugins/extruntime/extruntime.go — Runtime + Module interfaces, Manifest
- plugins/extruntime/manifest.go — JSON manifest loader (stdlib only)
- plugins/extruntime/native.go — reference impl, registers Go funcs

Concrete runtimes land in sibling commits:
- plugins/wasmvm — wazero (pure Go, hard sandbox)
- plugins/v8vm — v8go (cgo, -tags v8vm)
- plugins/gojavm — goja adapter (no new deps)
- plugins/extbench — fixtures + benchmarks across all 5

* feat(base): goja runtime adapter + extension loader

* feat(base): v8go extension runtime (cgo, -tags v8vm)

V8-via-cgo runtime satisfying plugins/extruntime.Runtime. Build-tag
gated so the default base binary never pulls v8go's ~50MB of cgo
deps — a stub at !v8vm reports the same Name/Capabilities and
errors at Load with ErrUnsupported.

JS calling convention: payload bytes are JSON.parse'd into a real
JS value and passed as the single argument to globalThis[fn]; the
return value is JSON.stringify'd back to bytes. Much simpler than
the wasm ptr/len dance — V8 has actual objects.

Single shared Isolate across all modules (creation is expensive,
isolates are not thread-safe so a global mutex serializes anyway).
Per-module Context pool (default 8, BASE_V8VM_POOL_SIZE) for state
isolation across back-to-back invocations. Heap cap intent via
BASE_V8VM_HEAP_MB_MAX; v8go v0.9.0 has no public per-isolate heap
API, so the value is advisory and enforcement is the V8 default
plus ctx-cancel TerminateExecution() watchdog.

Tests cover the cancel path (200ms timeout on `while(true)`), heap
bomb (2^30-char string), pool reuse cap, wrong-runtime manifest,
and closed-module rejection.

* feat(base): wazero extension runtime

Implements plugins/wasmvm — the wazero-backed extruntime.Runtime.

- Shared wazero.Runtime with compilation cache + WASI snapshot preview1
  instantiated once so guest modules can do stdio/clock/random.
- Per-Module instance pool (default 8, BASE_WASMVM_POOL_SIZE override)
  to amortize InstantiateModule across invocations.
- Host-guest ABI: guest exports fn(i32 ptr, i32 len) -> i64 (packed
  ptr<<32 | len), plus __base_alloc / __base_free for the buffer
  handshake. Missing alloc/free returns ErrUnsupported.
- Hard abort on ctx cancel via api.Module.Close(); cancelled instances
  are replaced from a background ctx so pool size stays stable.
- Tests cover identity, load+invoke, ctx cancel, concurrent pool
  reuse (16 invokes over pool size 4), invoke-after-close, and
  wrong-runtime manifest rejection. Fixture-dependent tests skip
  with TODO until extbench fixtures land.

* feat(base): extension benchmark fixtures across 5 runtimes

* feat(base): extension runtime benchmarks + writeup

plugins/extbench/ benches all five extension runtimes (native, goja,
wazero AS, wazero Javy, v8go) against an identical validate(email,age)
workload. Per-runtime Serial + Parallel + Coldstart + Memory tests.
run.sh runs the suite at -benchtime=2s -count=3 (~3 min total).

Headline numbers (Apple M1 Max, Go 1.26.3, median of 3):
  native     1,197 ns/op serial  /     694 ns/op parallel
  goja       4,513 ns/op serial  /   1,652 ns/op parallel
  wazero AS  9,956 ns/op serial  /   3,271 ns/op parallel
  v8go      11,895 ns/op serial  /  12,667 ns/op parallel (no parallelism)

Cold start ranges from native 17us to wazero AS 3.9ms. Binary size
delta from +89 KB (native) to +47.8 MB (all four with -tags v8vm).
Javy is wired but skipped pending a stdio-shim or custom Javy plugin
(its WASI command ABI doesn't match wasmvm's pointer-based contract).

Full writeup in docs/EXTENSIONS_BENCHMARK.md, raw runs in
plugins/extbench/bench-results.txt.

* feat(base): scale study — concurrent module / tenant / goroutine fanout

Five-section fan-out study answering:
1) per-loaded-module marginal memory (N up to 10000)
2) per-tenant concurrent invocation latency (T up to 1000) + pool-size elbow
3) sustained concurrency on ONE module (M up to 10000)
4) goroutine cost (1000 parked goroutines, HeapInuse delta per runtime)
5) cold pool startup vs steady-state (N=100 sequential Load)

Headlines: goja is the multi-tenant winner (~9KB/tenant, 188k ops/s at
M=1000); wazero's per-module cost is the 8-instance pool — elbow at
pool=4; v8go fatal-aborts above ~10 concurrent invokes (darwin/arm64
SIGSEGV inside libv8); native pays nothing per goroutine. Wazero
pool=256 explodes RSS to 10.8 GB at 1000 modules without latency win.

Implementation: cloneTenants() materializes a temp dir per tenant with
a unique extension name, symlinked module artifact, and native-handler
re-registration. v8go capped at concurrency=10 and skipped in
goroutine-cost — the cap itself is the data, recorded in the writeup.

Exports Validate from fixtures/native-go so scale_test can register
the same handler under N unique extension names without re-implementing.

Wall clock: <1 minute for the full study on Apple M1 Max.

* feat(base): pyvm runtime — cgo CPython 3.13 (-tags pyvm)

Fifth extension runtime alongside native/goja/wazero/v8go. Default
build links the stub; real impl behind `-tags pyvm` requires
pkg-config python-3.13-embed.

Architecture: one process-wide CPython init, per-module pool of PEP
684 OWN_GIL sub-interpreters (default 4 / BASE_PYVM_POOL_SIZE).
Each Invoke pins to OS thread, attaches a sub-interpreter, runs
__main__.fn(json.loads(payload)), returns json.dumps(result).
Free-threading (PEP 703) detected at runtime via Py_GIL_DISABLED.

Honest limitations documented in README + capability flags:
- HardSandbox=false (C extension segfault kills the host)
- SupportsAbort=false (no primitive to abort a running sub-interp
  from another OS thread without holding its GIL)

Direct cgo bridge (~150 lines), no DataDog/go-python3 or cpy3
dependency — both fail to compile against CPython 3.12+ because
they reference removed PyFloat_ClearFreeList /
PyDict_ClearFreeList. JSON marshaling stays inside Python's
stdlib json module so the cgo surface stays narrow.

Tests: 7 table tests pass on darwin/arm64 + Python 3.13.13.

* feat(base): pyvm fixture + bench integration

Adds plugins/extbench/fixtures/pyvm-py/ (validate.py + manifest)
implementing the same validate(email, age) spec as the other four
fixtures — pure stdlib, no third-party Python deps.

bench_test.go: BenchmarkPyvm_Serial + BenchmarkPyvm_Parallel skip
gracefully when built without -tags pyvm (same shape as v8go).

scale_test.go: pyvm registered alongside the four existing runtimes
with a 100-module cap (each module holds up to 4 sub-interpreters
at ~4.6 MB each — 1000 modules would be ~18 GB and dominate the
study). BASE_PYVM_POOL_SIZE plumbed through setPoolEnv /
restorePoolEnv for the pool-elbow subtest.

* docs: pyvm benchmark numbers vs existing 4 runtimes

Five-way throughput table on darwin/arm64 + Apple M1 Max + Python
3.13.13 (GIL-enabled). Headline:

- pyvm Serial 4,089 ns/op — between native (1,229) and goja (4,646)
- pyvm Parallel 2,765 ns/op — faster than wazero (4,530) and v8go
  (12,311), close to goja (1,982)
- pyvm B/op 496, allocs/op 10 — smallest footprint of any runtime
- pyvm cold start 14.3 ms/module — worst in the field; warm-load
  at boot, never lazy-load
- pyvm per-module RSS asymptote 4.6 MB — highest in the field

HIP-0105 verdict: experimental-by-default, production-recommended
for single-tenant runtime:pyvm deployments only. Multi-tenant must
add crash-isolation:none + hard-abort:none labels so operators
can policy-reject.

Free-threading (PEP 703) NOT measured — homebrew does not ship
[email protected] on darwin/arm64 as of 2026-05-18. Runtime detection
path is wired; verification needs a Linux host.

Bench crash mode found + fixed: defer-order LIFO race in
pyvm_leave/release on Parallel — restructured to straight-line
ordering. No further crash modes observed across 90+ s of load.

* feat(base): starlark extension runtime (plugins/starkvm)

Adds Google's go.starlark.net interpreter as the sixth extruntime
Runtime. Pure Go, no cgo, no GIL, per-thread Cancel() for cooperative
abort. Same Module shape as gojavm/wasmvm: pre-warmed pool, JSON payload
in / JSON result out, ctx-cancel honored.

Manifest: "runtime": "starlark", "module": "validate.star".

Fixture plugins/extbench/fixtures/starkvm-star/ implements the shared
validate(email, age) workload. Uses the permissive Starlark dialect
(while + recursion enabled via FileOptions); strict Bazel Starlark
would forbid while, doc note in the .star file.

Wired into bench_test.go (Serial/Parallel), coldstart_test.go,
memory_test.go (TestMemory_Starkvm), and scale_test.go (spec entry +
BASE_STARKVM_POOL_SIZE pool-elbow sweep). Tests cover capabilities,
roundtrip, unknown fn, post-close, ctx cancel mid-loop, pool reuse,
and the full validate fixture against 5 input cases (8 tests total).

First measurement on Apple M1 Max: 3701 B/invoke, 1 GC per 1000 calls
-- comparable to goja (3433 B/invoke).

* feat(extbench): wasm-Rust fixture + bench

Pure-Rust validate(email, age) compiled to wasm32-unknown-unknown,
loaded by the existing wasmvm runtime. Implements HIP-0105 pointer/
length ABI directly (no zip-rs dep, no Hanzo-internal deps). Compares
apples-to-apples with the AssemblyScript fixture: same ABI, same
payload, different source language.

- Cargo.toml: standalone crate (workspace=[] to escape hanzo workspace)
- src/lib.rs: no_std + bump allocator + tolerant JSON scrape
- build.sh: cargo build --release --target wasm32-unknown-unknown
- validate.wasm: 9.1 KB committed artifact (no toolchain dep at bench run)
- bench_test.go: BenchmarkWazero_Rust_{Serial,Parallel}

Run: ~4.5us serial / 3.9us parallel on M1 Max.

* docs: deferred CPython-WASI + RustPython ABI shim

Both runtimes emit WASI command modules (stdin/stdout via _start) which
don't match Base's wasmvm pointer ABI (validate(ptr,len) -> i64).

CPython-WASI deferred: requires either a 30 MB python.wasm artifact +
a stdio-piping Rust/AS shim, or building CPython from source against
the WASI SDK (~15-20 min) and forking _start. Out of 60-min budget.

RustPython deferred: tried embedding rustpython-vm 0.5.0 in a Rust
shim that exposes the pointer ABI directly. The {freeze-stdlib +
rustpython-compiler} feature combo fails to compile on wasm32-wasip1
because of cfg-gated paths in define_exception_fn!. Resolving requires
either patching the crate or waiting on upstream 0.6.x. Documented
the exact attempt + failure in fixtures/wazero-rustpython/README.md so
the next agent doesn't repeat the work.

Both fixtures stay in tree with extension.json + validate.py + README;
the wasmvm runtime reports "module not built" and the bench harness
Skip()s them with a clear message.

* test(extbench): add pyvm + wasm-Rust memory + coldstart tests

memory_test.go gains TestMemory_Pyvm + TestMemory_Wazero_Rust +
deferred-fixture skips (TestMemory_Wazero_CPythonWASI/RustPython).
runMemory helper now Skip()s when manifest.module artifact is
missing — same loadFixture path the bench harness uses, so
deferred fixtures (extension.json present, .wasm not built) skip
cleanly instead of failing.

coldstart_test.go gains BenchmarkColdstart_Pyvm + ..._Wazero_Rust +
the two deferred-fixture skips. benchColdstart already probes Load
and skips on failure, so the new ones inherit clean Skip behavior.

All tests pass under default + v8vm + pyvm tag combinations.

* bench: full cross-runtime comparison results

Captured 2026-05-19 on Apple M1 Max, default + -tags 'v8vm pyvm'.

8 runtimes measured (5 built, 3 deferred-by-ABI):
- native-go, goja, starkvm, wazero-as, wazero-rust, v8go, pyvm: full
  throughput + memory + cold-start numbers
- wazero-cpython-wasi, wazero-rustpython: Skip with "module not built"
  message (deferred per fixture README)

Headline:
- pyvm 6.9us serial (only native-go is faster) but 34ms cold-start
- starkvm 21us serial — comparable to goja, pure-Go, no JIT
- wazero-rust 14.6us serial — slightly better than wazero-as
- v8go is the slowest non-deferred runtime (cgo marshalling cost)

* docs: consolidated runtime comparison in EXTENSIONS_BENCHMARK.md

New section "Full Cross-Runtime Comparison (2026-05-19)" appended.

Covers 8 runtimes side-by-side:
- 5 originally measured (native, goja, wazero-as, v8go, pyvm)
- 3 added in this sprint (starkvm, wazero-rust, plus two
  documented deferrals: wazero-cpython-wasi + wazero-rustpython)

Sections:
- Throughput table (serial / parallel / B/op / allocs/op / coldstart)
- Per-invocation memory (TestMemory)
- Cold-start (BenchmarkColdstart)
- Scale + crash characteristics (qualitative)
- Recommendation grid (use case -> best/second/avoid)
- Headline findings (5 bullets, honest, including the deferral wall)

Numbers come from plugins/extbench/bench-results-all.txt; reproduce
section at the bottom.

* docs(pyvm): add measured 3.13t free-threading numbers + cpy3 fork note

- EXTENSIONS_BENCHMARK.md "Free-threading vs sub-interpreter
  parallelism" section now has the measured numbers (3.13 vs 3.13t,
  Apple M1 Max, default-bridge pyvm): free-threading is a non-event
  for the embedding workload (4,073 / 5,733 ns/op default-GIL vs
  4,478 / 5,805 ns/op free-threaded). PEP 684 OWN_GIL pool already
  delivers per-OS-thread parallelism; PEP 703 doesn't add headroom
  in this contract.
- Library-choice paragraph now references hanzoai/cpy3 (the
  canonical Hanzo Go binding for embedded CPython — fork of
  go-python/cpy3 with 3.12+ compat + sub-interp wrapper +
  free-threading detection + JSON hot-path helper + 3.13t build
  script) and explains why pyvm itself keeps the consolidated
  direct cgo bridge (one cgo call per Invoke, ~30% faster than
  routing through cpy3 wrappers for this workload).
- go.mod tidied — cpy3 no longer in pyvm's import graph; cpy3 is
  documented but not directly used by base.
2026-05-18 23:38:22 -07:00

213 lines
6.1 KiB
Go

package gojavm
import (
"context"
"encoding/json"
"errors"
"fmt"
"os"
"path/filepath"
"sync"
"github.com/dop251/goja"
"github.com/hanzoai/base/plugins/extruntime"
)
// interruptValue is what we pass to goja.Runtime.Interrupt() when ctx
// cancels mid-invocation. goja re-throws this as an interruptedError
// from RunString/Call; we detect it and translate back to ctx.Err().
type interruptValue struct{}
func (interruptValue) Error() string { return "gojavm: context canceled" }
// module is one loaded extension. The JS source is compiled to a
// *goja.Program once at Load. Per Invoke we borrow a runtime from the
// pool, ensure the program has been run on that runtime (registers
// globalThis.fn / globalThis.<export>), then call the requested
// function.
type module struct {
name string
exports []string
program *goja.Program
pool *vmsPool
// runtimeInit tracks which goja.Runtime instances have already had
// the program executed on them. Each pool slot only needs to run
// the script once; subsequent invocations just call globalThis.fn.
// Keyed by pointer identity. Cheap and correct.
mu sync.Mutex
runtimeInit map[*goja.Runtime]bool
closed bool
}
func newModule(pool *vmsPool, dir string, m *extruntime.Manifest) (*module, error) {
src := m.Module
if src == "" {
src = "index.js"
}
path := filepath.Join(dir, src)
data, err := os.ReadFile(path)
if err != nil {
return nil, fmt.Errorf("gojavm: read %s: %w", path, err)
}
// Compile once. goja.Compile is deterministic and the resulting
// *goja.Program is safe to share across runtimes.
prog, err := goja.Compile(path, string(data), true)
if err != nil {
return nil, fmt.Errorf("gojavm: compile %s: %w", path, err)
}
return &module{
name: m.Name,
exports: m.Exports,
program: prog,
pool: pool,
runtimeInit: map[*goja.Runtime]bool{},
}, nil
}
func (m *module) Name() string { return m.name }
func (m *module) Runtime() string { return "goja" }
func (m *module) Exports() []string { return m.exports }
func (m *module) Invoke(ctx context.Context, fn string, payload []byte) ([]byte, error) {
m.mu.Lock()
if m.closed {
m.mu.Unlock()
return nil, extruntime.ErrClosed
}
m.mu.Unlock()
if err := ctx.Err(); err != nil {
return nil, err
}
var out []byte
runErr := m.pool.run(func(vm *goja.Runtime) error {
// Ensure the module's program has been executed on this
// particular runtime, populating globalThis.<fn>.
if err := m.ensureLoaded(vm); err != nil {
return err
}
callable, ok := goja.AssertFunction(vm.Get(fn))
if !ok {
return fmt.Errorf("%w: %s:%s", extruntime.ErrUnknownFn, m.name, fn)
}
// Unmarshal the payload to a Go value; let goja convert it to
// a JS value. The runtime layer treats payloads as JSON bytes
// so every backend (wasm/v8/native/goja) sees the same wire.
var arg any
if len(payload) > 0 {
if err := json.Unmarshal(payload, &arg); err != nil {
return fmt.Errorf("gojavm: payload not JSON: %w", err)
}
}
// Watchdog: when ctx cancels, interrupt the runtime. The
// interrupted Call returns an *InterruptedError wrapping our
// interruptValue. Stop the watchdog once the call returns so
// we don't leak the goroutine.
done := make(chan struct{})
defer close(done)
go func() {
select {
case <-ctx.Done():
vm.Interrupt(interruptValue{})
case <-done:
}
}()
jsArg := vm.ToValue(arg)
result, err := callable(goja.Undefined(), jsArg)
if err != nil {
// Translate an interrupt back to the ctx error.
var iex *goja.InterruptedError
if errors.As(err, &iex) {
if ctxErr := ctx.Err(); ctxErr != nil {
return ctxErr
}
}
return fmt.Errorf("gojavm: invoke %s:%s: %w", m.name, fn, err)
}
// Marshal the result back to JSON bytes. goja's
// json.Stringify handles JS-native edge cases (undefined ->
// no field, functions skipped, Date -> ISO string) more
// faithfully than Go's encoding/json on the exported value.
out = stringify(vm, result)
return nil
})
if runErr != nil {
return nil, runErr
}
return out, nil
}
func (m *module) Close() error {
m.mu.Lock()
m.closed = true
m.runtimeInit = nil
m.mu.Unlock()
return nil
}
// ensureLoaded runs the compiled program on this runtime exactly once.
// Subsequent calls are no-ops. We track the runtime pointer identity;
// when a pool slot is recycled (it isn't in our pool, but if a future
// implementation rotates slots) the entry would simply be re-run.
func (m *module) ensureLoaded(vm *goja.Runtime) error {
m.mu.Lock()
already := m.runtimeInit[vm]
if !already {
m.runtimeInit[vm] = true
}
m.mu.Unlock()
if already {
return nil
}
if _, err := vm.RunProgram(m.program); err != nil {
// Roll back the init marker so a transient error doesn't
// permanently poison the slot for this module.
m.mu.Lock()
delete(m.runtimeInit, vm)
m.mu.Unlock()
return fmt.Errorf("gojavm: load %s: %w", m.name, err)
}
return nil
}
// stringify reproduces `JSON.stringify(value)` semantics in Go without
// re-entering the runtime. goja.Value.Export gives us a Go-side view;
// json.Marshal on that view produces the canonical wire bytes.
//
// We prefer this to `vm.RunString("JSON.stringify(...)")` because:
// - it avoids parsing a literal expression each call,
// - it avoids polluting globalThis with a temp var, and
// - undefined results become an empty []byte rather than the
// literal string "undefined" (which is not valid JSON).
func stringify(vm *goja.Runtime, v goja.Value) []byte {
if v == nil || goja.IsUndefined(v) || goja.IsNull(v) {
// `null` is valid JSON; for parity with JSON.stringify(null) -> "null"
// we emit literal null. Undefined -> empty bytes (the spec says
// JSON.stringify(undefined) returns undefined, not a string).
if v != nil && goja.IsNull(v) {
return []byte("null")
}
return nil
}
exported := v.Export()
b, err := json.Marshal(exported)
if err != nil {
// Fall back to the runtime's own stringifier, which handles
// quirky values (Symbol, BigInt, cycles) by throwing — we
// turn that throw into an empty result.
return nil
}
return b
}