on this page

linux-drivers.com

dataset

Evidence-backed dossiers for 864 Linux kernel drivers, recommending which legacy drivers to keep, annotate, deprecate, or remove

period: 2026-present
tech:
Linux KernelStatic AnalysisCodexJSON SchemaAstro

linux-drivers.com is an evidence-backed map of which legacy drivers in the Linux kernel deserve a closer look for deprecation or removal — and which should stay. It publishes 2,028 per-driver dossiers, each with a recommendation, a confidence score, and citations linking every non-trivial fact to a real source (kernel.org, lore.kernel.org, vendor EOL notices, distro package indexes, OpenWrt and postmarketOS wikis, etc.).

How it was built

A two-phase pipeline:

  1. Phase 1 — git-log dormancy ranker (no LLM, ~3 minutes). Walks drivers/, computes a dormancy score from raw commit activity minus mechanical churn (treewide sweeps, SPDX tags, bot commits). Filters out non-driver content and active mega-subsystems (amdgpu, i915, xe, iwlwifi, etc.). Produces a ranked top-N shortlist.
  2. Phase 2 — codex CLI dossier probe (~75s per driver). For each shortlisted driver, runs an openai/codex exec with a strict JSON Schema, --output-schema-enforced output, an MCP-backed lore.kernel.org search tool, and web search. The model returns a single JSON dossier matching the schema.

A third validation step checks that every cited URL resolves, every dossier validates against the schema, and recommended deprecations align with concrete evidence.

What’s in the corpus

verdictcountmeaning
remove12active upstream removal patch series exists
deprecate52hardware obsolete, no active maintenance, replacement exists
keep-annotate603mostly inactive but plausible niche use; document caveats rather than remove
keep813active or strategically important; no action needed
not-a-driver548content directory misclassified by phase-1

The list is intentionally conservative: only ~3% of probed drivers were flagged for any change, and remove was reserved for cases where someone is already proposing the patch upstream.

Audience

  • Subsystem maintainers deciding what to deprecate next
  • Kernel developers writing removal patches who want pre-built evidence chains
  • Distro packagers auditing what to drop from their kernel configs
  • Researchers studying kernel longevity

The data and dossiers are CC-BY-4.0; the scripts are MIT.

on this page