1Decision
The zram swap capacity of the Pixel targets is raised from the stock 50% to 100% of RAM (about 7.3 GiB) by setting mmd.zram.size=100% through PRODUCT_PRODUCT_PROPERTIES in config/mk/google_devices/common/device-common.mk of vendor/adevtool, and removing the stock mmd.zram.size entries from config/device/common/sysprop-inclusion.yml, as the patch package memory/zram-size in sovereignos-platform.
It does not touch secure app spawning, ART, the zygote, lmkd or any hardening default; it is the first of the memory levers recorded in stallion-memory-pressure-investigation.md, which also keeps the others (fewer processes started at boot, a smaller per-process boot image copy) open.
2Why and evidence
Owner decision 2026-10-10 (chat): accepted with MEM-011 still NOT_RUN; the effect on kills and the plateau of SwapFree are measured on stallion after a user candidate. Owner direction 2026-10-08 (chat): optimise RAM, CPU and swap without any security regression, measure first.
Measured on stallion 2026100701: with spawning on the swap has 58 MB free at 178 s after boot (0 MB at 422 s, anon demand 6.3 GB) against 1227 MB free with it off (3.3 GB); the physical zram was 813 MB for the full 3.7 GB, so the slot cap is the limit and not RAM. On lynx the override made SwapTotal 6112248 kB, with the physical zram unchanged (661 MB against 631 and 673 MB).
Source: post_process_props.py fails a second assignment with another value, so the stock line has to be removed (tested with the real tool).
3Consequences
The raise is a limit, not a reservation; a full 7.3 GiB swap would hold about 1.6 GB of physical zram at the measured 4.5 times.
The images up to 2026101002 set mmd.zram.writeback.limit.enabled=false, so mmd applies no per-run writeback limit (writeback.rs, update_writeback_limit returns at once and the result of calculate_idle_writeback_limit is unused); the 1 GiB backing file is the only bound there, and whether the larger capacity writes more to flash is not measured (MEM-005).
The images on tag 2026100600 set it to true with mmd.zram.writeback.max_bytes_per_day=25769803776 (24 GiB a day) Measured · build.prop of lynx 2026110204, so there the daily cap bounds the writes, not the file. The stock lines for bluejay and oriole (3221225472) go too; those devices are not built. The generated product.prop files of the build server carry the stock line and need the same removal before the build (SOURCE.md).
The effect on lmkd kills is measured on lynx only as headroom (3189 MB free against 887 to 1135 MB; lynx never ran out), and not yet on stallion, a user build that refuses the DeviceConfig flag (mmd_native/zram.size is not on the allowlist). Built into lynx image 2026100801 and checked there (SwapTotal 6112248 kB from build.prop, no crash entries); not built for stallion. The first value was 80% (patch 0001, the lynx image above).
On 2026-10-09 the owner chose 100% on the condition that it could be verified to be better: lmkd treats free swap under 10% of the total as low (lmkd.cpp swap_free_low_percentage), so 90% of the capacity is usable, and in a fixed-load experiment on lynx (MEM-006, one run per size) 80% sat under that line 55% of the time and 100% never, with 1405 MB against 1452 MB of physical zram.
Patch 0002 is landed on the build server (c0fa4e99) and in the package; built into lynx image 2026101001 with ADR-151 (MEM-012) and flashed there (MEM-013: SwapTotal 7640312 kB against MemTotal 7640316 kB). What is not verified: the effect on kills (lynx has too few cached apps) and how much swap stallion needs over a day; the plateau of SwapFree there decides whether 100% is enough (MEM-004).
4Review when
The built image shows two mmd.zram.size lines or a value other than 100%; kills do not drop on stallion; the swap-in latency shows in the daily use; an upstream mmd change of the property name or the sysprop filter.
Read from docs/governance/decision-log.md, entry ADR-150.