Perf/stress ng gap 17

[hodgesds]

Summary

• Propagate typed ReservePressure allocation failures through NUMA policy paths, eliminating duplicate aggregate free-memory snapshots from anonymous demand faults.

• Restrict aggregate allocator statistics to published online NUMA nodes.

• Add Linux-style placement for ordinary private anonymous mmap:

  • Honors valid aligned address hints.
  • Reuses gaps created by munmap.
  • Falls back to bottom-up first-fit placement.
  • Checks both regular and huge-page VMAs.
  • Serializes placement and publication under the existing VMA transaction lock.

• Preserve existing MAP_FIXED, eager-population, protection, zeroing, and reclaim semantics.

• Update the memory specification, tests, and TCB safety argument.

Performance

Directional release-build results for:

stress-ng --vm 1 --vm-method cache-lines --vm-bytes 256M --timeout 10s --verify

• NARF median: 12,758.27 bogo-ops/s
• Linux median: 11,963.26 bogo-ops/s
• NARF/Linux: 106.6%

These are three controlled one-vCPU KVM samples, not the formal N≥30 performance protocol.

Validation

• cargo fmt --all -- --check
• cargo clippy --all-targets -- -D warnings
• Host tests
• x86_64 release memory tests: 364 passed
• aarch64 release memory tests: 282 passed
• x86_64 and aarch64 release syscall ABI suites
• Three verified release stress-ng runs
• Safety-argument validation
• git diff --check