gc: Perry-owned memory for GC arena blocks (2 MiB OS regions, huge pages by complete units)

[proggeramlug]

Perry now owns the memory under its GC heap on Linux. Arena blocks come straight from the OS in 2 MiB-aligned regions instead of from the Rust global allocator (mimalloc), and page size is chosen by how much of a 2 MiB unit is used, not by object lifetime.

What changes

• OS regions for arena blocks. Each agent maps 2 MiB-aligned regions and pools its blocks. mimalloc keeps ordinary native allocations only. Other platforms keep std::alloc and the 1 MiB geometry.
• 2 MiB GC blocks on Linux. One usable block per aligned region, so a block and a huge page are the same unit.
• One page-size rule. Every generation gets MADV_HUGEPAGE on complete 2 MiB units, and sub-unit extents and rounded tails stay on base pages. A nursery is refilled every cycle, so it is dense, and it gets huge pages like mature memory. There is no prefault, no collapse request and no tuning constant.
• The defrag selector decides from the allocation kind stored in the block's existing metadata (it fits in padding), not from a size threshold.
• Census bitmap capacity follows the block size.
• Separate fix in the GC trigger: a parse bump compared against the nursery-capped effective trigger, so it could lower an already armed whole-arena threshold. With bigger blocks this re-armed a minor on almost every small parse (fastify: 379 minors instead of 17). Now the bump compares with the whole-arena threshold itself, and it never lowers it.

What it removes: GC memory no longer goes through a general allocator that can't know GC lifetimes or density, and the page-size decision is one rule instead of a nursery/mature split.

Measurements

These were measured on c85c2db (before #12103), on qb6 CPUs 56–63 under the shared measurement lock. n=5 interleaved runs, default auto-optimized builds, and every output matches Node. The rebase onto current main only crossed Buffer/stream test files.

Program	RSS MiB main → this	sys ms main → this	wall s main → this
tsc	206.1 → 176.9	23 → 32 (noise 49)	1.010 → 1.022
fastify inject	262.2 → 226.5	86 → 77 (noise 26)	11.63 → 11.75
Effect	181.4 → 153.1	26 → 35 (noise 17)	2.324 → 2.306
Zod ×5000	48.8 → 32.6		1.129 → 1.128
qs parse	50.9 → 34.5		2.116 → 2.109
qs stringify	52.8 → 36.9		4.533 → 4.530
commander	46.6 → 30.2		0.745 → 0.743
buffer_heavy	91.6 → 73.2		0.800 → 0.834
worker_heavy	209.9 → 116.0		0.205 → 0.218
upm lock	211.7 → 148.6		0.453 → 0.445
upm cold	361.8 → 251.7		0.540 → 0.627
hello	1.6 → 1.6

• System CPU on tsc, fastify and Effect is within the same-binary noise of main. Wall medians are within the host's noise, and no speedup is claimed.
• Most RSS saving comes from not backing sparse GC memory with huge pages inside mimalloc. A THP-off control on main reaches similar RSS, but costs page faults; this change keeps huge pages where memory is dense.
• An earlier variant with base pages for the nursery saved about 5 MiB more on small programs (Zod, qs, commander) but cost +40–60 ms system CPU on tsc, fastify and Effect.

Known limits:

• Some programs show instruction differences of about ±0.1% against main that are not fully attributed. Part of it comes from the measured binaries' path length changing startup allocation.
• A whole nursery unit gets huge pages even when a small program only partly fills it. That costs about 5–6 MiB on small programs, and an advise-after-fill step would be a follow-up.

Tests

• cargo test --release -p perry-runtime -- --test-threads=1: 5206 passed, 0 failed. Plain main fails only the known z8 RSS flake (Expire retired receiver shapes in the class read cache #12111). The build is clean under -D warnings.
• The z8 stream churn test now restores the payload-prototype, runtime-handle, shape-rekey, promise and timer scanners after its copying-GC isolation guard. Without that, the guard dropped roots the workload uses and the test crashed. Its assertions are unchanged.
• Five parity subsets pass with none skipped.
• New tests: region mapping and alignment, census capacity at both geometries, the allocation-kind defrag selector, page advice (complete units huge, tails base), a fastify-like churn that keeps the minor count equal at 1 MiB and 2 MiB, and worker transfer with sender exit.
• Each was checked by reverting its fix: the census capacity, the non-moving extent filter, the pool charge, the promotion advice, nursery advice and the trigger comparison. Every test fails without its fix.

Open question for review

A diagnostic control (z8 without the isolation guard, at the larger batch count) hit the young-log verifier at gc/young_log.rs:173: young log for shapes.families+indices does not name …, reached through scan_shape_table_rekey_mut during a moving minor. Main at a matched larger count passed that control, so this is not shown to be pre-existing. It does not reproduce with the guard, and the full suite passes. Reviewers who own the shape table: please check whether the larger blocks expose a missing young-log entry there. The backtrace is available on request.

[coderabbitai]

Note

Currently processing new changes in this PR. This may take a few minutes, please wait...

⚙️ Run configuration

• Configuration used: defaults
• Review profile: CHILL
• Plan: Advanced
• Run ID: 58db139c-e9e1-41e7-a1df-3197d2c3e4d8

📥 Commits

Reviewing files that changed from the base of the PR and between 68e41fa and 7961e37.

📒 Files selected for processing (29)

• changelog.d/gc-region-manager-2mib.md
• crates/perry-runtime/src/arena/block.rs
• crates/perry-runtime/src/arena/block/reuse_window_tests.rs
• crates/perry-runtime/src/arena/mod.rs
• crates/perry-runtime/src/arena/page_meta/mod.rs
• crates/perry-runtime/src/arena/promote.rs
• crates/perry-runtime/src/arena/quarantine.rs
• crates/perry-runtime/src/arena/region.rs
• crates/perry-runtime/src/arena/tests.rs
• crates/perry-runtime/src/ffi/perry_memory_profile.c
• crates/perry-runtime/src/gc/large_extent_selection_tests.rs
• crates/perry-runtime/src/gc/oldgen_defrag.rs
• crates/perry-runtime/src/gc/policy.rs
• crates/perry-runtime/src/gc/tests/adopt_census.rs
• crates/perry-runtime/src/gc/tests/census_block_windows.rs
• crates/perry-runtime/src/gc/tests/census_two_mib.rs
• crates/perry-runtime/src/gc/tests/debt_pacer.rs
• crates/perry-runtime/src/gc/tests/evacuation.rs
• crates/perry-runtime/src/gc/tests/incremental_sweep_reclaim.rs
• crates/perry-runtime/src/gc/tests/mod.rs
• crates/perry-runtime/src/gc/tests/native_payload_streams.rs
• crates/perry-runtime/src/gc/tests/oldgen.rs
• crates/perry-runtime/src/gc/tests/runtime_roots.rs
• crates/perry-runtime/src/gc/tests/sweep_window_births.rs
• crates/perry-runtime/src/gc/tests/triggers.rs
• crates/perry-runtime/src/gc/trace.rs
• crates/perry-runtime/src/startup_memory_profile.rs
• test-files/_helpers/region_transfer_sender_exit.ts
• test-files/test_gap_region_transfer_sender_exit.ts

 ____________________________________________________________________________________________
< If you don't start with a spec, every piece of code you write is a patch. - Leslie Lamport >
 --------------------------------------------------------------------------------------------
  \
   \   (\__/)
       (•ㅅ•)
       / 　 づ

✨ Finishing Touches

📝 Generate docstrings

• Commit to this branch
• Create a new PR

🧪 Generate unit tests (beta)

• Commit to this branch
• Create a new PR

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}