JavaScriptCore's Garbage Collector (Riptide)

Riptide is non-moving, generational, parallel, mostly-concurrent, conservative-on-the-stack. Understanding those five words prevents most GC bugs in Bun.

The mental model

The heap is a graph. GC does a breadth-first search from roots → marks everything it reaches → everything unmarked is freed (lazily, on next allocation from that block). It does NOT compact or move objects — pointers stay stable for an object's lifetime.

Two collection modes:

• Eden GC: only scans newly-allocated objects + remembered set. Fast, frequent.
• Full GC: scans everything. Slower, rarer.

It runs concurrently. Marking happens on background threads _while JS is executing_; the mutator only stops at brief safepoints. visitChildren runs off the main thread, racing with your code.

How the VM gathers roots

Roots are not a hardcoded list — they are marking constraints registered with Heap::addMarkingConstraint() and run to fixpoint. The built-in set lives in Heap::addCoreConstraints() (vendor/WebKit/Source/JavaScriptCore/heap/Heap.cpp:2970):

Bun registers an additional constraint, DOMGCOutputConstraint (src/jsc/bindings/BunGCOutputConstraint.cpp), which calls visitOutputConstraints on every marked cell in Bun's output-constraint subspaces (event targets, generated classes with visitAdditionalChildren, etc.).

Constraint volatility controls when they re-run during the fixpoint:

• GreyedByExecution — may produce new grey cells whenever the mutator runs (re-run after every resume)
• GreyedByMarking — may produce new grey cells when _other_ marking happens (re-run after each drain)
• SeldomGreyed — usually doesn't add anything; run last

Object layout: the 8-byte JSCell header

Every GC-managed object inherits JSCell (runtime/JSCell.h):

| StructureID (4) | indexingTypeAndMisc (1) | JSType (1) | flags (1) | cellState (1) |

• StructureID — compressed hidden-class pointer
• indexingTypeAndMisc — 2 bits are an embedded WTF::Lock (the cell lock); always CAS this byte
• cellState — inlined GC color, used by the write barrier

Out-of-line, in the MarkedBlock footer (or PreciseAllocation header for objects >~8KB):

• isMarked bit — survived last GC
• isNewlyAllocated bit — allocated since last GC

Liveness = isMarked || isNewlyAllocated (with logical-versioning so blocks aren't swept eagerly).

CellState and the write barrier

vendor/WebKit/Source/JavaScriptCore/heap/CellState.h:

PossiblyBlack   = 0   // visited (or old-space-pending-rescan during full GC)
DefinitelyWhite = 1   // new / unmarked
PossiblyGrey    = 2   // on the mark stack

Generational + concurrent GC share one retreating-wavefront barrier:

// After: obj->field = newValue
if (obj->cellState <= blackThreshold)   // 0 normally, bumped while GC is marking
    writeBarrierSlowPath(obj);          // → put obj on remembered set / revisit

You almost never write this by hand. Use WriteBarrier<T> as the field type and call .set(vm, owner, value) — it stores then barriers. A raw JSCell* / JSValue member without a WriteBarrier wrapper is a bug: eden GC will free the target out from under you.

LazyProperty<Owner, T>, LazyClassStructure, and WriteBarrierStructureID are barrier-aware variants for lazily-initialized fields and structures.

Allocation: where objects live

bmalloc/libpas provides pages; JSC carves them up:

• MarkedBlock — 16KB block, fixed cell size (segregated free list). Footer holds bitvectors. 16-byte minimum cell alignment. addr & ~(16KB-1) → block, so liveness checks are O(1).
• PreciseAllocation — large objects (>~8KB), individually malloc'd, 96-byte GC header. Always returns addresses with addr % 16 == 8 so ptr & 8 distinguishes them from MarkedBlock cells.
• CompleteSubspace — size-segregated set of BlockDirectorys for general JS objects.
• IsoSubspace — one subspace per C++ type (security: a freed cell can only be reused for the _same_ type, defeating type-confusion UAF). Every Bun class with native fields needs its own IsoSubspace — subspaceFor<T> in the header, slot in BunClientData/DOMIsoSubspaces.

Allocation may trigger GC. A safepoint exists at every allocation. Never assume "I just allocated X, so Y from before is still alive" unless Y is rooted.

Conservative stack scanning — what it does and doesn't guarantee

vendor/WebKit/Source/JavaScriptCore/heap/ConservativeRoots.cpp walks the native stack/registers word-by-word (after MachineThreads::tryCopyOtherThreadStacks snapshots them). Any aligned word inside a live MarkedBlock cell or PreciseAllocation is a root.

This means: a JSCell* / JSValue in a C++/Rust local variable _usually_ keeps the object alive — no Handle/Local ceremony like V8.

This does NOT mean you're always safe. The compiler may dead-store-eliminate the local after its last visible use, or never spill it. If you extract an interior pointer (string->characters8(), butterfly storage, typed-array vector()) and then call something that can allocate, the original cell may no longer be on the stack:

JSC::EnsureStillAliveScope keepAlive(cell);   // RAII: forces cell onto stack until scope end
// ... use interior pointer, call things that allocate ...

or ensureStillAliveHere(cell). In Rust: value.ensure_still_alive().

visitChildren — the per-cell tracing hook

// In header:
DECLARE_VISIT_CHILDREN;
WriteBarrier<JSObject> m_callback;
WriteBarrier<Unknown>  m_cachedValue;

// In .cpp:
template<typename Visitor>
void JSFoo::visitChildrenImpl(JSCell* cell, Visitor& visitor) {
    auto* thisObject = jsCast<JSFoo*>(cell);
    ASSERT_GC_OBJECT_INHERITS(thisObject, info());
    Base::visitChildren(thisObject, visitor);   // ALWAYS call base first

    visitor.append(thisObject->m_callback);
    visitor.append(thisObject->m_cachedValue);
}
DEFINE_VISIT_CHILDREN(JSFoo);

Rules — runs concurrently on a GC thread:

• No allocation. No toJS, no jsString, nothing that touches vm.heap.
• No ref()/deref() of RefCounted (not thread-safe).
• No locks the main thread might also take while allocating (deadlock).
• If a field can be torn by a racing mutator, take Locker locker { thisObject->cellLock() } in both visitChildren and the mutating site.
• Forgetting to append() a WriteBarrier field → use-after-free, often eden-GC-only, often only under load.

visitAdditionalChildren and output constraints

visitChildren only sees the cell's own fields. When a JS wrapper's liveness should propagate to other JS objects reachable through native state (event listeners, observers, the JS values held inside a wrapped C++ object), Bun uses the WebCore pattern:

// Custom hook called from BOTH places:
template<typename Visitor>
void JSFoo::visitAdditionalChildren(Visitor& visitor) {
    wrapped().listeners().visitJSEventListeners(visitor);
    visitor.addOpaqueRoot(&wrapped());
}

// 1) From visitChildren (normal marking):
DEFINE_VISIT_CHILDREN_WITH_MODIFIER(..., JSFoo) {
    ...
    thisObject->visitAdditionalChildren(visitor);
}

// 2) From visitOutputConstraints (constraint fixpoint re-scan):
template<typename Visitor>
void JSFoo::visitOutputConstraints(JSCell* cell, Visitor& visitor) {
    auto* thisObject = jsCast<JSFoo*>(cell);
    Base::visitOutputConstraints(thisObject, visitor);
    thisObject->visitAdditionalChildren(visitor);
}

Why two entry points? visitChildren runs once when the cell turns grey. But marking may later discover that some _other_ native object (an opaque root) is live, which retroactively makes more of _this_ cell's references live. visitOutputConstraints is re-invoked by DOMGCOutputConstraint during the constraint fixpoint to catch that.

To make a class participate, its IsoSubspace must be registered as an output-constraint subspace (clientSubspaceFor* with outputConstraint in BunClientData / generated ZigGeneratedClasses.cpp). The codegen does this automatically when .classes.ts has hasPendingActivity, own properties, or event-target semantics.

Opaque roots — liveness through non-JSCell pointers

When native objects form a graph that should keep wrappers alive:

// In some wrapper's visitAdditionalChildren:
visitor.addOpaqueRoot(nativePtr);          // "nativePtr is reachable"

// Elsewhere, deciding whether ANOTHER wrapper survives:
bool JSBarOwner::isReachableFromOpaqueRoots(Handle<Unknown> h, void* ctx,
                                            AbstractSlotVisitor& v, ASCIILiteral* reason) {
    auto* bar = static_cast<Bar*>(ctx);
    if (UNLIKELY(reason)) *reason = "Bar is in document tree"_s;
    return v.containsOpaqueRoot(bar->ownerNode());
}

The opaque-root set is just a HashSet<void*> rebuilt each cycle. It's how DOM trees stay alive as a unit.

JSC::Weak<T>, WeakImpl, WeakBlock, WeakHandleOwner

JSC::Weak<T> (vendor/WebKit/Source/JavaScriptCore/heap/Weak.h) is the GC-aware weak pointer. It does not keep its target alive; .get() returns nullptr after the target is collected.

Under the hood:

• Each Weak<T> owns a WeakImpl* (vendor/WebKit/Source/JavaScriptCore/heap/WeakImpl.h): { JSValue, WeakHandleOwner* (low bits = state), void* context }. State is Live → Dead → Finalized → Deallocated.
• WeakImpls are slab-allocated in 1KB WeakBlocks (vendor/WebKit/Source/JavaScriptCore/heap/WeakBlock.h, blockSize = 1024). Every MarkedBlock and PreciseAllocation has a WeakSet — a linked list of WeakBlocks for cells in that container.
• During the Ws constraint, each WeakBlock::visit() walks its WeakImpls; for each one whose target is not yet marked, it calls WeakHandleOwner::isReachableFromOpaqueRoots(handle, context, visitor, &reason). Return true → the target is marked (the weak ref is "upgraded" this cycle). This is how hasPendingActivity() and opaque-root reachability keep wrappers alive even when nothing strongly references them.
• After marking, WeakBlock::reap() flips unmarked Live impls to Dead. WeakBlock::sweep() later runs WeakHandleOwner::finalize(handle, context) on each Dead impl, then frees the slot. finalize runs on the mutator thread but the cell is already dead — do not touch its JS fields. Typical use: drop the wrapper from a native→JS wrapper cache.

struct MyOwner final : public JSC::WeakHandleOwner {
    bool isReachableFromOpaqueRoots(Handle<Unknown>, void* ctx,
                                    AbstractSlotVisitor& v, ASCIILiteral*) override {
        return static_cast<NativeThing*>(ctx)->hasPendingActivity();
    }
    void finalize(Handle<Unknown>, void* ctx) override {
        static_cast<NativeThing*>(ctx)->m_wrapper = nullptr;
    }
};
JSC::Weak<JSFoo> m_wrapper { jsFoo, &myOwnerSingleton, nativeThing };

Weak<T> is move-only (allocates a WeakImpl). Don't put it in a hot path; cache it.

JSRef — the native↔wrapper reference pattern

When a native object needs to hold a reference back to its own JS wrapper, use JSRef (src/jsc/JSRef.rs), not gcProtect, not a raw JSValue field, and usually not Strong directly.

JSRef is a tagged union with three states:

• Weak — a bare JSValue. Does not keep the wrapper alive. Valid only because the wrapper's finalize() will flip this to Finalized before the cell is freed, so try_get() returns None instead of a dangling pointer. (This is _not_ a JSC::Weak; it's cheaper — no WeakImpl allocation.)
• Strong — wraps bun_jsc::Strong (a JSC::Strong<Unknown> root). Keeps the wrapper alive.
• Finalized — terminal; try_get() returns None.

Pattern: strong while busy, weak while idle.

this_value: JSRef, // initialized with JSRef::empty()

// On construction / when work starts:
self.this_value.set_strong(js_wrapper, global);   // or .upgrade(global)

// When the last in-flight operation completes:
self.this_value.downgrade();                      // Strong -> Weak, GC may now collect

// In any callback that needs the wrapper:
let Some(js_this) = self.this_value.try_get() else { return };

// In the codegen'd finalize():
self.this_value.finalize();

See ServerWebSocket, UDPSocket, MySQLConnection, ValkeyClient for real examples.

JSRef requires a finalizer. The Weak state is only sound because the codegen'd finalize() flips it to Finalized before the cell is reused. If your .classes.ts entry has finalize: true (almost all native-backed classes do), JSRef is the default choice for self-references.

JSRef vs hasPendingActivity: prefer JSRef. hasPendingActivity: true is a GC-thread-polled atomic predicate; its only real justification is when many concurrent operations independently keep the wrapper alive and there's no single place to call upgrade()/downgrade() — i.e., refcount-style liveness where the count is touched from multiple threads. That's uncommon. If you can identify "work started" / "work finished" edges, use JSRef. Don't add hasPendingActivity reflexively; it costs a constraint-fixpoint poll on every GC.

gcProtect / JSValueProtect — almost never

gcProtect() / JSValueProtect() push into Heap::m_protectedValues (a ref-counted root map, visited by the Msr constraint). It's the legacy C-API mechanism. Avoid it in Bun:

• It's a raw global root with manual unprotect — easy to leak.
• It has no owner, so heap snapshots can't attribute the retention.
• bun_jsc::Strong / JSRef give the same guarantee with RAII and a destructor.

The only legitimate uses are inside the JSC C API shims themselves, or one-off debugging.

Extra-memory reporting — reportExtraMemoryAllocated / reportExtraMemoryVisited

The GC schedules itself by bytes-allocated-since-last-GC. It only sees JSCell allocations, so a 32-byte wrapper around a 50MB native buffer looks like 32 bytes → GC never triggers → OOM.

Contract — both halves are required:

// 1) When the native memory is allocated (or the wrapper takes ownership):
vm.heap.reportExtraMemoryAllocated(ownerCell, byteCount);

// 2) In visitChildren, every time the cell is visited:
visitor.reportExtraMemoryVisited(thisObject->wrapped().byteSize());

• reportExtraMemoryAllocated adds to the "since last GC" counter and may immediately trigger a GC (it's a safepoint). Call it _after_ the cell is fully constructed.
• reportExtraMemoryVisited adds to the "live bytes after this GC" counter, which sets the next trigger threshold. If you forget this half, the heap's high-water mark drifts down each cycle and you get back-to-back full GCs (the "GC death spiral").
• If the size changes over time, report the delta on growth (reportExtraMemoryAllocated(cell, newSize - oldSize)) and report the current size in visitChildren.
• deprecatedReportExtraMemory exists for callers that can't satisfy the visit-side half — avoid it.

In .classes.ts, estimatedSize: true generates the reportExtraMemoryVisited side; you implement estimated_size() in Rust. You still call reportExtraMemoryAllocated (or the binding's helper) at allocation time.

HeapAnalyzer — heap snapshots and labelling

vendor/WebKit/Source/JavaScriptCore/heap/HeapAnalyzer.h is the abstract visitor used to build heap snapshots (Web Inspector "Heap Snapshot", and Bun's V8-compatible BunV8HeapSnapshotBuilder). When a snapshot is requested, marking runs with an analyzer attached and each cell's analyzeHeap static is called:

void JSFoo::analyzeHeap(JSCell* cell, HeapAnalyzer& analyzer) {
    auto* thisObject = jsCast<JSFoo*>(cell);
    Base::analyzeHeap(cell, analyzer);
    analyzer.setWrappedObjectForCell(cell, &thisObject->wrapped());
    analyzer.setLabelForCell(cell, thisObject->wrapped().url().string());
    if (auto* child = thisObject->m_callback.get())
        analyzer.analyzePropertyNameEdge(cell, child, vm.propertyNames->callback.impl());
}

API (HeapAnalyzer):

• analyzeNode(cell) — record a node
• analyzeEdge(from, to, RootMarkReason) / analyzePropertyNameEdge / analyzeVariableNameEdge / analyzeIndexEdge — record a labelled edge
• setWrappedObjectForCell(cell, void*) — link wrapper → native pointer
• setLabelForCell(cell, String) — human-readable name in the snapshot
• setOpaqueRootReachabilityReasonForCell — why a weakly-held wrapper survived

If your class shows up as an opaque blob in heap snapshots, implement analyzeHeap.

How to keep things alive (decision table)

Destruction & finalizers

• static constexpr bool needsDestruction = true → C++ destructor runs when the cell is swept. Sweep is lazy (next allocation from that block, or IncrementalSweeper), so destruction is delayed arbitrarily. Do not rely on it for prompt resource release — expose explicit close()/dispose().
• In .classes.ts, finalize: true → native finalize() called from the destructor. Same laziness applies.
• WeakHandleOwner::finalize runs earlier (at weak-reap time) but the cell is already dead; only use it to clear caches.
• Destructors run on the mutator thread but other JS objects may already be swept — do not dereference WriteBarrier fields in a destructor.

Debugging GC issues

# Force synchronous, frequent GC — turns rare races into immediate crashes
BUN_JSC_collectContinuously=1 BUN_JSC_useConcurrentGC=0 bun-debug test.js

# Zero free cells so UAF reads are obvious
BUN_JSC_scribbleFreeCells=1

# Validate the GC's own bookkeeping
BUN_JSC_verifyGC=1 BUN_JSC_verboseVerifyGC=1

# See what's being collected / heap growth
BUN_JSC_logGC=2 BUN_JSC_showObjectStatistics=1

# Force GC from JS
Bun.gc(true)      // sync full GC
require('bun:jsc').heapStats()

If a bug only reproduces with concurrent GC on → missing write barrier or visitChildren race. If it only reproduces with collectContinuously=1 → something isn't rooted across an allocation. If memory grows but heapStats().heapSize doesn't → missing reportExtraMemoryAllocated. If GC runs constantly with little garbage → missing reportExtraMemoryVisited.

Key source files

• vendor/WebKit/Source/JavaScriptCore/heap/Heap.cpp — collectImpl, addCoreConstraints (root list)
• vendor/WebKit/Source/JavaScriptCore/heap/SlotVisitor.cpp / SlotVisitorInlines.h — drain(), append, addOpaqueRoot, reportExtraMemoryVisited
• vendor/WebKit/Source/JavaScriptCore/heap/MarkedBlock.h, vendor/WebKit/Source/JavaScriptCore/heap/PreciseAllocation.h — cell containers, isLive
• vendor/WebKit/Source/JavaScriptCore/heap/CellState.h, runtime/WriteBarrier.h, runtime/WriteBarrierInlines.h
• vendor/WebKit/Source/JavaScriptCore/heap/ConservativeRoots.cpp, vendor/WebKit/Source/JavaScriptCore/heap/MachineStackMarker.cpp
• vendor/WebKit/Source/JavaScriptCore/heap/Weak.h, vendor/WebKit/Source/JavaScriptCore/heap/WeakImpl.h, vendor/WebKit/Source/JavaScriptCore/heap/WeakBlock.h, vendor/WebKit/Source/JavaScriptCore/heap/WeakSet.h, vendor/WebKit/Source/JavaScriptCore/heap/WeakHandleOwner.h
• vendor/WebKit/Source/JavaScriptCore/heap/HeapAnalyzer.h, vendor/WebKit/Source/JavaScriptCore/heap/HeapSnapshotBuilder.cpp, Bun: vendor/WebKit/Source/JavaScriptCore/heap/BunV8HeapSnapshotBuilder.cpp
• vendor/WebKit/Source/JavaScriptCore/heap/DeferGC.h, vendor/WebKit/Source/JavaScriptCore/heap/Strong.h, vendor/WebKit/Source/JavaScriptCore/heap/HandleSet.h
• runtime/JSCell.h / JSCellInlines.h — header layout, visitChildren base
• Bun: src/jsc/bindings/BunGCOutputConstraint.cpp, ZigGeneratedClasses.cpp (codegen'd visitChildren / visitOutputConstraints)