feat(memory): apply kernel memory tunables before memory-mode runs

[not-matthias]

Benchmark repos running codspeed --mode memory currently need a hand-written CI step to stabilise kernel memory behaviour before measuring: disable THP, zero vm.compaction_proactiveness / vm.swappiness / kernel.numa_balancing, swapoff -a, and drop the page cache before each suite.

The runner now does this itself at the start of every memory-mode execution.

Behaviour

• Applied only when a CI run environment is detected, so developer machines are not silently mutated. Skipped entirely when privileges cannot be elevated without a password prompt.
• Best effort: a knob that cannot be applied is a warning, never fatal.
• MemoryTunables::apply() returns a guard holding the previous value of every knob it changed, and restores it on drop — including on the error paths, where teardown() never runs. Only knobs that were not already at the target are captured, and THP is restored to the exact mode it held (madvise stays madvise).
• swapoff -a is guarded: skipped on zram devices (it permanently resets their disksize) and whenever the swapped pages would not fit in MemAvailable. Restoring re-enables the recorded entries one by one, since swapon -a would miss a swap file absent from /etc/fstab.
• The page cache drop has no previous state to restore; the node is a write-only trigger.
• Applied in run() rather than setup(), so the page cache is dropped before every suite and --skip-setup does not bypass it. The sysctl/THP writes are no-ops on later suites since ensure_sysctl compares before writing.

Commits

1. refactor(executor): move linux_sysctl into executor helpers — pure move, the sysctl primitive is no longer walltime-specific.
2. feat(memory): apply kernel memory tunables before memory-mode runs

Verification

• cargo clippy --all-targets clean, cargo test --lib (pre-existing failures on this host are the sudo-requiring tests, which write kernel.kptr_restrict and are identical on main).
• Local non-CI run logs Not running in CI, skipping kernel memory tunables; host THP unchanged.
• Privileged container: THP enabled/defrag become [never], vm.compaction_proactiveness = 0, vm.swappiness = 0, and the swap guard correctly logged Leaving swap enabled: swapped pages do not fit in available memory.

[codspeed-hq]

Merging this PR will not alter performance

⚠️ Unknown Walltime execution environment detected

Using the Walltime instrument on standard Hosted Runners will lead to inconsistent data.

For the most accurate results, we recommend using CodSpeed Macro Runners: bare-metal machines fine-tuned for performance measurement consistency.

✅ 17 untouched benchmarks

---

_{Comparing cod-3300-kernel-memory-tunables (fc2fe71) with main (8868af5)}

[greptile-apps]

Greptile Summary

The PR applies and restores Linux kernel memory tunables around memory-mode benchmark runs and centralizes reusable sysctl guards for profiling. The current restoration path does not preserve swap priorities, while privileged local runs remain able to mutate host memory state.

• Adds best-effort THP, swap, and page-cache stabilization around memory runs.
• Records changed THP and swap state in a drop guard.
• Moves profiling sysctl handling into shared executor helpers and restores prior values.

Confidence Score: 3/5

The PR is not yet safe to merge because privileged local runs mutate host memory state and swap restoration changes non-default priorities.

Memory tuning is reached locally whenever root or passwordless sudo is available, including an irreversible page-cache drop, and restored swap areas lose their configured priority because only their paths are recorded.

Files Needing Attention: src/executor/memory/tunables.rs and src/executor/memory/executor.rs

Important Files Changed

Filename	Overview
src/executor/memory/tunables.rs	Adds host memory tuning and restoration, but swap restoration loses each area's original priority and the existing local-run admission defect remains.
src/executor/memory/executor.rs	Applies the tuning guard for every memory execution, including privileged local runs.
src/executor/helpers/linux_sysctl.rs	Centralizes sysctl mutation behind a guard that restores changed values on drop.
src/executor/wall_time/executor.rs	Retains profiling sysctl guards on the executor so prior values survive setup and are restored when the executor is dropped.

Sequence Diagram

sequenceDiagram
  participant E as MemoryExecutor
  participant T as MemoryTunables
  participant K as Linux kernel
  participant B as Benchmark
  E->>T: apply()
  T->>K: Set THP and disable eligible swap
  T->>K: Drop page cache
  E->>B: Run memory benchmark
  B-->>E: Exit
  E->>T: Drop guard
  T->>K: swapon(path)
  Note over T,K: Original swap priority is not restored
  T->>K: Restore THP mode

Loading

_{Reviews (3): Last reviewed commit: "fix(walltime): restore profiling sysctls" | Re-trigger Greptile}

[art049]

I know it was the case before but it hurts readability a bit

[art049]

Is it possible to gate this whole mod behind the feature flag but not every single line, because now we end up gating almost every single line with #[cfg(target_os = "linux")] 😅

[not-matthias]

@greptileai review again