[codex] Fix parallel mark claiming

src/heap.zigmodified+19-6

Changes to src/heap.zig

		⋯436 linesShow the 436 lines above this hunk@@ -437,7 +437,7 @@ pub fn Heap(comptime Binding: type) type {
437	437	/// and a mutator's `writeBarrier` never both push the same cell; the
438	438	/// single-threaded path is a plain check.
439	439	fn claimMark(self: *Self, h: *Header) bool {
440		- if (self.concurrent.load(.acquire)) {
	440	+ if (self.parallel or self.concurrent.load(.acquire)) {
441	441	if (@cmpxchgStrong(bool, &h.marked, false, true, .acq_rel, .monotonic) != null) return false;
442	442	_ = @atomicRmw(usize, &self.marked_count, .Add, 1, .monotonic);
443	443	return true;
		⋯105 linesShow the 105 lines above this hunk@@ -549,10 +549,20 @@ pub fn Heap(comptime Binding: type) type {
549	549	/// stopped, so:
550	550	/// - The whiten pass + state reset run under `alloc_lock`, the same
551	551	/// leaf lock `create` takes, so the O(n) `all`-list walk and the
552		- /// `marked` clears can't race a peer's prepend / born-grey mark-bit
553		- /// set. `alloc_lock` is a leaf (never held across a safepoint or a
554		- /// per-structure lock), so guarding the walk with it cannot deadlock
555		- /// a mutator that is parked or spinning for an object lock.
	552	+ /// born-grey mark-bit set can't race a peer's prepend. `alloc_lock`
	553	+ /// is a leaf (never held across a safepoint or a per-structure
	554	+ /// lock), so guarding the walk with it cannot deadlock a mutator
	555	+ /// that is parked or spinning for an object lock. The whiten *stores*
	556	+ /// to `marked` are atomic because `alloc_lock` does NOT serialize the
	557	+ /// barrier path: a peer still finishing a store from the *previous*
	558	+ /// cycle (it read `marking` true before the prior finish cleared it)
	559	+ /// reaches `claimMark`, whose CAS atomically touches `marked` without
	560	+ /// `alloc_lock`. That CAS is always a benign no-op here — its target
	561	+ /// was reachable-and-marked last cycle, so the strong CAS fails and
	562	+ /// `writeBarrier` returns before mutating any list (a successful CAS
	563	+ /// would mean a swept-garbage target, which the terminal root
	564	+ /// handshake already rules out). Atomic-vs-atomic is race-free; the
	565	+ /// new cycle's happens-before is the `marking`=true release below.
556	566	/// - `marking`+`concurrent` are published *while still holding*
557	567	/// `alloc_lock`, so no cell can be prepended between "whitened" and
558	568	/// "barrier armed": a `create` that wins the lock after us already
		⋯12 linesShow the 12 lines above this hunk@@ -571,7 +581,10 @@ pub fn Heap(comptime Binding: type) type {
571	581	std.debug.assert(self.parallel);
572	582	self.lockAlloc();
573	583	var it = self.all;
574		- while (it) |h| : (it = h.next) h.marked = false;
	584	+ // Atomic store: a lagging peer's `claimMark` CAS (barrier path, no
	585	+ // `alloc_lock`) can touch the same `marked` byte concurrently. See
	586	+ // the whiten note in the doc comment above.
	587	+ while (it) |h| : (it = h.next) @atomicStore(bool, &h.marked, false, .monotonic);
575	588	self.marked_count = 0;
576	589	self.mark_stack.clearRetainingCapacity();
577	590	self.weak_slots.clearRetainingCapacity();