Repository navigation
[perf] Minor GC does O(old-gen) work: full-region sweep walk, whole-heap RS rebuild, per-cycle page-meta iterations #6181
Description
Problem
Minor GC cost includes several O(old-gen) components even when nothing old is touched:
- Sweep visits/rewrites every old-gen header on every minor (
gc/oldgen.rs:1181-1225cursor over all regions; flags/accounting rewritesgc/oldgen.rs:1291-1311,1338,1357; minors setreclaim_dead_old_blocks=falsebut still walk,gc/cycle.rs:1390-1416). - 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 withrequire_marked=falsefor minors it re-dirties pages from unreachable-but-unswept old parents. - 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).
Activity
perf(gc): minor collections skip the whole-heap old→young RS rebuild (#…
proggeramlug commentedon Jul 10, 2026
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 commentedon Jul 10, 2026
Per-sub-cost status on current main (post the July 9-10 GC campaign):
gc/cycle.rs:1405-1429,// Fix 2 (#6181)) — minors early-exit pastRememberedSetRebuild; regression teststest_minor_skips_whole_heap_old_to_young_rebuild+test_minor_preserves_old_to_young_edge_across_minorscover it. Don't re-touch.ArenaSweepObjectsState::newalways builds aBlockIndexcursor 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.old_pages_begin_gc_cycle/old_pages_reset_sweep_accounting/old_page_summarywalk 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 unfilteredarena_walk_objects_with_block_indexjust to filter by active block range — a filtered helperarena_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-phaseValidPointerSet, a different structure — verified by diff.)Recommendation: land the
policy.rs:765filtered-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. Thefix/6083-nursery-scoped-minorbranch (#6129) is closed/superseded by #6239 and not reusable here.perf(gc): use the filtered arena walk for the minor promotable-surviv…
proggeramlug commentedon Jul 11, 2026
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 vacuously0 == 0).Status of this issue's three sub-costs:
policy.rs:765unfiltered-walk half is now fixed (perf(gc): use the filtered arena walk for the minor promotable-survivor census (#6181) #6277). Theold_pages_*reset-accounting walks remain, coupled to (a).Leaving open for (a) + the coupled reset-accounting.
perf(gc): lazy epoch reset for old-page dirty-slot accounting (#6181) (…
proggeramlug commentedon Aug 9, 2026
Closing — two of the three sub-costs are fixed, and the third is refiled so it is visible as its own work.
Fixed: (b) whole-heap remembered-set rebuild and (c) per-cycle page-meta iterations, via #6213 / #6277 / #6568.
Refiled: #7716 — sub-cost (a), the minor's full-region sweep walk over the old generation (
gc/oldgen.rs::sweep, still on the minor path at HEAD), which makes minor cost scale with old-gen size rather than nursery survivors.That issue carries the two traps that make it less mechanical than it looks: the malloc sweep is a separate, trigger-gated path that must not be conflated with it, and #7682/#7690's precedent that removing work from the minor path can livelock on its own, because the nursery cap was keyed on bytes the changed collector no longer frees.