[perf] Minor GC does O(old-gen) work: full-region sweep walk, whole-heap RS rebuild, per-cycle page-meta iterations

[proggeramlug]

Problem

Minor GC cost includes several O(old-gen) components even when nothing old is touched:

1. Sweep visits/rewrites every old-gen header on every minor (gc/oldgen.rs:1181-1225 cursor over all regions; flags/accounting rewrites gc/oldgen.rs:1291-1311,1338,1357; minors set reclaim_dead_old_blocks=false but still walk, gc/cycle.rs:1390-1416).
2. The remembered-set finalize rebuild walks the ENTIRE heap (all arena objects + full malloc registry) every cycle (gc/cycle.rs:1244-1264, gc/verify.rs:279-349), and with require_marked=false for minors it re-dirties pages from unreachable-but-unswept old parents.
3. Old-page bookkeeping iterates every old 4 KB page meta up to 3× per minor (arena/page_meta.rs:555-571, gc/oldgen.rs:181-184,1199), plus the promotion-handoff pre-walk over all arena objects once survivor in-use ≥ 24 MB (gc/policy.rs:864-902).

Minor pause grows monotonically with process lifetime on servers.

Proposed fix

• Restrict the minor sweep cursor to nursery+survivor regions; maintain old-gen accounting incrementally from promotion/reclaim events; per-block live counts so fully-dead block reset needs no per-object walk.
• Maintain the RS incrementally (the barriers already do) and restrict the finalize rebuild to evacuated/moved parents — rebuild_evacuated_old_to_young_remembered_set (gc/verify.rs:243-253) plus the CI: gc_write_barrier_stress red on main (missing old→young remembered-set edges → segfault) #5029 snapshot repair already exist for exactly this.
• Lazy epoch-counter resets for page metas; keep defrag candidates on a small heap updated at promotion time.

Context

2026-07-09 GC audit (pause/scalability + barrier dimensions).

[proggeramlug]

Status: the whole-heap old→young RS rebuild for minors was eliminated in #6213 (minors jump AtomicFinalize→Sweep; barriers + evacuation-sticky + #5029 snapshot repair carry coverage). Remaining from this issue: the minor sweep's full-region walk and per-cycle census cost — the census half just landed for budgeted cycles in #6239 (no exact set built at all); the sweep-walk restriction to young regions remains open and is queued in the #6180 flip campaign.

[proggeramlug]

Per-sub-cost status on current main (post the July 9-10 GC campaign):

• (b) whole-heap remembered-set rebuild on every minor: FIXED by perf(gc): minor collections don't walk/rebuild over old-gen (#6181) #6213 (gc/cycle.rs:1405-1429, // Fix 2 (#6181)) — minors early-exit past RememberedSetRebuild; regression tests test_minor_skips_whole_heap_old_to_young_rebuild + test_minor_preserves_old_to_young_edge_across_minors cover it. Don't re-touch.
• (a) full old-gen sweep walk on every minor: still present, deliberately deferred. ArenaSweepObjectsState::new always builds a BlockIndex cursor over all 5 regions incl. Old (gc/cycle.rs:1337, arena/walk.rs). PR perf(gc): minor collections don't walk/rebuild over old-gen (#6181) #6213 scoped this out with a real correctness blocker: a minor's mark phase marks root/young-reachable old objects, and the minor sweep's full old-gen walk is the only place those MARKED bits get cleared before the next full GC (which assumes clean marks at entry). Skipping it naively → stale marks → next full GC skips tracing → frees live old children → UAF. Needs a design choice (minor-scoped mark suppression vs. a full-GC mark-clear pass + delta accounting) — not a quick patch.
• (c) per-cycle page-meta iteration: partly present. old_pages_begin_gc_cycle/old_pages_reset_sweep_accounting/old_page_summary walk all page metas every cycle (the reset-accounting one is coupled to (a)). One clean, decoupled win exists: copied_minor_promotable_active_survivor_bytes (gc/policy.rs:765-786) calls the unfiltered arena_walk_objects_with_block_index just to filter by active block range — a filtered helper arena_walk_objects_filtered(block_filter, cb) (arena/walk.rs:686-775) already exists and skips out-of-range blocks in O(n_blocks). It's a pure measurement pass, no mark-bit coupling → safe S swap. (This was not addressed by perf(gc): census-free budgeted cycles + RSS floor (#6179, #6180 Stage 2) #6239, whose "census" is the mark-phase ValidPointerSet, a different structure — verified by diff.)

Recommendation: land the policy.rs:765 filtered-walk swap as an isolated S/low-risk PR (essentially conflict-free with the #6180/#6243 churn); keep (a) + the coupled reset-accounting as a separate design-gated follow-up once the mark-clear approach is chosen. The fix/6083-nursery-scoped-minor branch (#6129) is closed/superseded by #6239 and not reusable here.

[proggeramlug]

Landed: #6277 — the promotable-survivor census (gc/policy.rs) now uses the block-filtered arena walk instead of the unfiltered one, so sub-cost (c)'s one cleanly-decoupled piece is done: the pure measurement pass drops from O(all arena objects) to O(active-survivor objects) + O(blocks). Regression test asserts the filtered census equals the old unfiltered-walk value byte-for-byte (with a guard that the equality isn't vacuously 0 == 0).

Status of this issue's three sub-costs:

• (b) whole-heap RS rebuild on every minor — fixed earlier by perf(gc): minor collections don't walk/rebuild over old-gen (#6181) #6213.
• (c) per-cycle page-meta iteration — the policy.rs:765 unfiltered-walk half is now fixed (perf(gc): use the filtered arena walk for the minor promotable-survivor census (#6181) #6277). The old_pages_* reset-accounting walks remain, coupled to (a).
• (a) full old-gen sweep walk on every minor — still open, still design-gated. A minor's mark phase marks root/young-reachable old objects, and the minor sweep's full old-gen walk is the only place those MARKED bits get cleared before the next full GC (which assumes clean marks at entry). Skipping it naively → stale marks → next full GC skips tracing → frees live old children (UAF). Needs the design choice (minor-scoped mark suppression vs. a full-GC mark-clear pass + delta accounting) before any patch.

Leaving open for (a) + the coupled reset-accounting.