Efuse programming: port the vendor PG-packet write path (append-only logical-map updates)

オープン
#285 コメント 0 件 リアクション 0 件 担当者 0 名 GitHub で見る

メンテナーはふだん 1 日以内に返信

まだ誰も着手していません。

評価

難易度
5/5
見積もり時間
1週間以上
初心者へのやさしさ
12/100
issue の種類
機能追加
明瞭さ
明確に書かれている
活発さ
静か
技術スタック
c
領域
embedded-iot

調査の方向性

Start with the existing read power sequence in src/jaguar1/EepromManager.cpp and the 8822E OTP path in src/jaguar3/HalJaguar3.cpp; compare them with the vendor write primitives and packet format. Implement the generation-neutral encoder in src/ with pseudo-test coverage before attempting hardware-specific writes. Done means safe dry-run and space accounting, verified packet writes with recovery, and family-specific hardware validation; burns require sacrificial adapters.

索引モデルが issue の本文から書いたものです。

説明

enhancement

Motivation

The vendor drivers ship a full efuse-programming stack (rtw_efuse_map_write → Efuse_PgPacketWrite → efuse_OneByteWrite, exposed as the MP-mode iwpriv efuse_set wmap,...). Devourer is read-only. Mirroring the write path lets us permanently fix MAC/board-option/ID fields on adapters — including resurrecting corrupt-efuse units — without carrying a vendor kernel module around.

Prerequisite: #284 (logical-map override). Its dump + patch grammar is the authoring format, and the RAM-only override is the mandatory pre-burn validation step: dump → edit → run with override → validate on-air → burn.

Write model (mirror of rtw_efuse_map_write)

Efuse is one-way: blank bytes read 0xFF, programming only pulls bits toward 0, nothing is ever erased. "Overwrite" = diff the desired logical map against the current one, encode only the changed 2-byte word-units as new PG packets, append them at the current end of the physical chain (Efuse_GetCurrentSize); later entries shadow earlier ones at materialization time.

Delivery: a dedicated efusetool example (dump / diff / dry-run / burn) — not wired into the demos. Dry-run is the default and prints the exact packet list + physical byte cost vs. remaining budget; burning requires an explicit flag, and MAC/VID/PID regions a second one.

Implementation details / hazards

  1. The write-mode power sequence is not the read one. Devourer's EfusePowerSwitch8812A (src/jaguar1/EepromManager.cpp:1322) only implements the read bracket; the vendor write path additionally raises programming voltage (LDO bits + REG_EFUSE_BURN_GNT = 0x69, the bWrite branch). Getting it wrong does not error — it produces silent partial burns. Port verbatim per chip.
  2. Per-generation write primitives diverge. Jaguar1: one-byte protocol on EFUSE_CTRL 0x0030. Jaguar2/3 vendor writes go through HalMAC (halmac_write_efuse), and our HalMAC port is read-only — decide between porting the halmac write entry points vs. raw-register writes with re-derived brackets. 8822E is a third primitive entirely: OTP power-cut + burst bracket, V1 register layout (RDY at bit 29, 11-bit address field) (src/jaguar3/HalJaguar3.cpp:522-573).
  3. 8822E burns must happen pre-InitWrite — the existing cache comment records that post-bring-up OTP access is unreliable (TX/coex interference). efusetool on Jaguar3-E is a dedicated cold session, never a live-device operation.
  4. Space accounting is chip-specific and unforgiving. Extended 2-byte headers past section 16 (Jaguar1/2), packed 2-byte headers everywhere on 8822E, and the vendor reserves an OOB-protect tail (EFUSE_OOB_PROTECT_BYTES) that must never be consumed. Overrun = chain tail lands in reserved space → undefined parse on next boot.
  5. badworden recovery is mandatory. Vendor verifies each word post-burn and re-programs failed words in a fresh packet; without it a marginal cell silently corrupts the value and wastes the space. Read-back verify + retry from day one.
  6. Interrupted-burn risk is real on USB. The header must precede data (the chain is positional), so an unplug mid-packet leaves a header pointing at 0xFF data and everything after may misparse. No clean mitigation exists: keep burns single-packet-small, re-parse and verify the full chain after each packet, document "don't burn through a flaky hub".
  7. Port the vendor efuse masks (per-chip protected-byte arrays covering calibration: xtal, thermal baseline, TX-power PG). Without them one typo'd offset permanently spends spare space shadowing a calibration byte.
  8. Testability without sacrificing hardware: port the vendor's bPseudoTest idea as a simulated efuse backend so diff→packet→space→extended-header logic runs as headless ctest selftests. Hardware validation still needs a sacrificial adapter — every burn, including successful tests, permanently spends bytes; budget the test plan honestly (a handful of burns per adapter lifetime).
  9. Shared logic placement: the diff→PG-packet encoder is generation-neutral pure code → src/ next to PhyTableLoader; only the byte-read/byte-write/power-bracket primitives live per-HAL.

Scope exclusions

  • BT efuse bank (8821A) — out.
  • Kestrel families — out.
  • PCIe 8821CE: likely works through RtlAdapter, but the is_usb() power-seq gates need checking before claiming it.

Suggested milestones

  1. Pseudo-test backend + PG-packet encoder + space accounting in src/, headless selftests (incl. extended-header and 8822E packed-header encodes).
  2. Jaguar1 write primitive (power bracket + one-byte write + read-back verify), validated on a sacrificial 8812AU: burn one scratch logical byte, cold-cycle, confirm shadowing.
  3. efusetool UX: dump / diff vs. #284 patch file / dry-run / burn, masks enforced.
  4. Jaguar2, then Jaguar3-8822C, then 8822E primitives — each gated on having a sacrificial adapter of that family.
主要言語
C
スター
56
フォーク
29
平均マージ
3時間 43分
マージ済み PR(30日)
75

環境構築

このプロジェクトには開発コンテナ、Dockerfile、コントリビューションガイドがありません。まず README を読み、一般的な手順ははじめてのコントリビューションガイドを参照してください。

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

OpenIPC/devourer のほかの issue

OpenIPC/devourer の issue をすべて見る

似ている issue

C の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。