v2.1.12-rc3(pre-release)——#146 描述符池诊断载具、#99 交棒重锁修复、GOP 顺序默认开、maker 一键安装预发布2026-11-12
v2.1.12-rc3(pre-release)——#146 描述符池诊断载具、#99 交棒重锁修复、GOP 顺序默认开、maker 一键安装
这是 pre-release;Latest 仍为 v2.1.11。 资产:BOOTX64.EFI.2.1.12-rc3(TX 构建事务版)、BOOTX64.EFI.readonly.2.1.12-rc3(只读负控版)、nv-uefi-unlocker-v2.1.12-rc3.zip(不含 ssdiag——杀软误报隔离政策,ssdiag 一律单独下载)、ssdiag-x86_64-windows-0.7.6.exe(独立资产)。校验和以 SHA256SUMS.2.1.12-rc3 为准。
本版有两个用途,请按现象对号
1. 需要 #146 现场数据时,它是诊断载具。 B550 winload 0xc000009a 的根因已于 2026-11-12 查明:winload 把交棒内存图的每一条描述符装进一个 1024 条静态池(winload.efi!BlpMmInitialize,48B/条,无动态扩容),池耗尽即 0xC000009A;现场分界 = 998 条能启动 / 1005 条失败,十个数据点零例外。本版在交棒前把整图按类型计数打行输出(见下),并对 >950 条打硬警告。相对 rc2 事务链零改动(铁律:只加只读采样)。
2. 被旧版的「EFI 全绿、进 Windows 后清零」打中过,它是修复载具。 本版含 061aef8:三代统一默认门开交棒。这类现象已确定机理(#99 留开等两台机器的回放),现场与判读见下一节——复测不要用 v2.1.11,它早于修复。
#99 清零型:已在本版修复,复测别用 v2.1.11
症状:EFI 事务全绿、读回全对;进 Windows 约 20 s 内 nvlddmkm 连发 id=14(事件原文 GPU recovery action changed from 0x0 (None) to 0x1 (PF FLR)),随后 FEAT 被清零、ssdiag 判 NOT_UNLOCKED。
触发者是我们自己的尾巴:旧版(≤v2.1.11)在交棒前无条件写 0x409650 ← 0xffffff8f 重锁 FEAT 门(硬件读回 0xffffffef)。Windows 驱动 init 随后把门推回 0xffffff8f,并发起一次 PF FLR——Turing 的 FEAT 块不活过 FLR,选择器覆盖随门一起被清掉。所以它看起来像「驱动代际问题」,实际与驱动版本无关。
修复:061aef8 默认门开交棒,卷根放一个 \tx-relock.txt 才恢复旧重锁。同一张 1f07、同 630.20 / GSP 630.03、同 OS 26340,只把载具从 v2.1.8 换成 v2.1.12-rc2(本版事务链与之相同,另加只读采样):交棒 0xffffffff、系统内仍 UNLOCKED、金丝雀 intact、卡与根口双 D0、0 条 nvlddmkm 事件;FP16 同频归一 1.71–1.77×。证据与整条判读链在 issue99-1f07-x99-gate-open,机制分类见 FAQ 机制表第 7 行。
版本下限(重要):061aef8 只进 v2.1.12-rc1 及以后。v2.1.11(tag a06aecf,11-08 14:26)在它之前 2 小时 44 分、仍无条件重锁,拿 v2.1.11「复测」只会得到假阴性。本页 rc3 已含该修复。
回报请带:同一 boot 的锁态基线(先跑一次 ssdiag)、金丝雀与 POWER: 段。SS0/SS1=0 本身不是清零证据——stock 基线会跨 boot 变(同卡两次 stock boot 实测 0x00040000/0x6 与 0x0/0x0),只有同一 boot 的前后可比才有判别力。
自 v2.1.12-rc2 以来的变化
1. CHAIN-MEMMAP-TYPES 直方图(新,只读):交棒摘要行之后按 EFI 内存类型输出全图描述符计数(reserved/loadercode/…/pmem/oem 16 桶)+ 小块碎片读数(tinyconv(<=64KiB):块数 + 合计 MiB)。判读:tiny-conv 多 = 可挤压碎片(disconnect 候选);MMIO/RESERVED 重 = 固件布局问题(只能 BIOS 修);LOADER/BSD 重 = 运行期驻留(驱动方压降)。
2. 描述符条数硬警告(新,只读):(... N descriptors ...) 总数 >950 时打 CHAIN-MEMMAP: WARNING 行,点名 winload 1024 池与断电重试(BIOS 版本对照协议见下)。950 线按现场最紧门校准(bootmgfw 同款池但 working set 仅几条,winload 倒欠 ~20–25 条)。
3. #228 GOP 句柄顺序修复默认开启:多卡 + 无头 BMC 拓扑下,bootmgr/grub 取第一个 GOP 句柄,而 native-GOP 重建把控制台卡的新子句柄追加到队尾 → 引导阶段黑屏。修复在 native-GOP 控制台重建通过验证之后,把其余无 EDID 卡的驱动停启一次,让它们的子句柄让到队尾、控制台子句柄排第一。四道门全过才动手(控制台子句柄在场 + 它带 EDID + 它当前不在首位 + 其余卡都无 EDID),任一失败 = 大声日志 + 零写入,术后读回;拓扑两义(竞争者带 EDID)时日志点名改用 \tx-gop-hide.txt。健康机负控 = Noop(零写入)。范围仅本次链内(解锁→交棒→OS),热重启的固件侧重排不在范围内;探针/翻转/隐藏旋钮可关(docs/csm-compat.md)。只动显示路径,不碰事务链。
4. 幻键防护三轮落地(#220/PR#222,J0s3f):确认键下沉等待(INPUT-SETTLE)+ 确认后按键抑制表 + \tx-ignore-key.txt 免重建过滤;Esc 永远豁免(abort 通道不可抑制)。
5. maker 一键安装提议 + 资产瘦身(make-usbroot.ps1 / .sh 双语同步):制盘完成后探测可用内置/外置目标 FAT32 分区并对内置盘提议免 U 盘 FAT32 安装;对外置盘提议 U 盘制盘安装(显式确认、可 -NoInstallPrompt / -InstallTo 跳过或指定);产物树瘦身。
6. 工具与诊断(非 loader):tools/disable-rtd3.sh(Linux RTD3 掉锁一键修复,FAQ#linux 流程自动化);ssdiag 0.7.4→0.7.6(GA100VFIO 词典扫盲、镜像探测、GR topology 锚点全家桶、X99 双路 inventory 两处现场捕获修复)——zip 内已不含 ssdiag。
rc1/rc2 已含(跳版用户速览)
1. 三代交棒默认门开(061aef8,FEAT 交棒前不再重锁;\tx-relock.txt 恢复旧行为)——rc2 说明第 3 条的同一件事,本文上方 #99 一节是它第一个由丢失现场翻绿的证据。
2. STAGING-FALLBACK 自动 LOW 带重试(rc1)+ rc2 的 FLR→FWSEC 重燃→LOW 重建载具修正(#156/#74/#191)。
3. \tx-preload.txt opt-in 化(#198,默认无预载)。
其余整表见 rc2 release notes。
#146 参与者:请用 rc3 回报
用 rc3 跑一次(老盘直接换 EFI\BOOT\BOOTX64.EFI,marker 照旧),把 loader.log 里两行发回来:
``
CHAIN-MEMMAP: ... (N descriptors, ...)
CHAIN-MEMMAP-TYPES: counted=N/N reserved=... | tiny_conv(<=64KiB): ...
`
N 就是判决数(>\~1000 = 条数顶板,断电重试/BIOS 对照;≤\~1000 且 largest block 腰斩 = 尺寸顶板)。BIOS 版本对照协议(Q-Flash 三轮、每版 2–3 boot 只报 N)见 上一条评论。
---
v2.1.12-rc3 (pre-release) — #146 descriptor-pool diagnostics vehicle, #99 handoff-relock fix, GOP ordering default-ON, maker one-click install
Pre-release; Latest remains v2.1.11. Assets: BOOTX64.EFI.2.1.12-rc3 (TX transaction build), BOOTX64.EFI.readonly.2.1.12-rc3 (read-only control), nv-uefi-unlocker-v2.1.12-rc3.zip (no ssdiag inside — AV-false-positive isolation policy; ssdiag is always a separate download), ssdiag-x86_64-windows-0.7.6.exe (standalone). Verify against SHA256SUMS.2.1.12-rc3.
Two purposes — match yours to the symptom
1. For #146 field data, this is the diagnostic vehicle. The B550 winload 0xc000009a root cause was pinned 2026-11-12: winload ingests every descriptor of the handoff memory map into a static pool of exactly 1024 entries (winload.efi!BlpMmInitialize, 48-byte stride, no dynamic overflow); pool exhaustion = 0xC000009A. Field line: 998 descriptors boot, 1005 fail — zero exceptions in ten data points. This build prints a per-type histogram of the whole map before handoff (below) and a hard warning above 950 entries. Zero transaction-chain changes versus rc2 (iron rule: read-only sampling only).
2. If the old build hit you with "EFI all green, wiped inside Windows", this is the fix carrier. It contains 061aef8: all three families now hand off with the gate OPEN by default. The shape is concluded (#99 stays open pending replays on two machines); the field symptom and how to read it are in the next section — do not retest on v2.1.11, it predates the fix.
#99 zeroing shape: fixed in this build, do not retest on v2.1.11
Symptom: the EFI transaction is all green with perfect readbacks; within ~20 s of Windows, nvlddmkm fires id=14 repeatedly (event text GPU recovery action changed from 0x0 (None) to 0x1 (PF FLR)), then the FEAT block is zeroed and ssdiag says NOT_UNLOCKED.
The trigger was our own tail: builds ≤ v2.1.11 unconditionally wrote 0x409650 ← 0xffffff8f (relock) before handoff (reads back 0xffffffef). The Windows driver init then pushes the gate back to 0xffffff8f and raises one PF FLR — the Turing FEAT block does not survive an FLR, so the selector overrides die with the gate. It looks like a driver-generation problem; it is not.
Fix: 061aef8 hands the gate off open by default; only a \tx-relock.txt on the volume root restores the legacy relock. Same 1f07, same 630.20 / GSP 630.03, same OS 26340, changing only the carrier (v2.1.8 → v2.1.12-rc2; this build runs the same transaction chain plus read-only sampling): handoff 0xffffffff, still UNLOCKED in-OS, canary intact, card and root port both D0, zero nvlddmkm events — and FP16 1.71–1.77× normalized to equal clocks. Evidence and the full judgment chain: issue99-1f07-x99-gate-open; mechanism class: FAQ mechanism row 7.
Version floor (important): 061aef8 exists only in v2.1.12-rc1 and later. v2.1.11 (tag a06aecf, 11-08 14:26) is 2 h 44 min older and still relocks unconditionally — "retesting" on v2.1.11 only yields a false negative. This rc3 build already carries the fix.
Please report: a same-boot locked baseline (run ssdiag first), plus the canary and POWER: sections. SS0/SS1=0` alone is not proof of a wipe — the stock baseline varies across boots (two stock boots on the same c
v2.1.12-rc2 — STAGING-FALLBACK retry vehicle fixed (FLR + FWSEC re-ignition)预发布2026-11-09
v2.1.12-rc2(pre-release)——STAGING-FALLBACK 重试载具修正(FLR + FWSEC 重新点火)
这是 pre-release;Latest 仍为 v2.1.11。 本版是 #156(1e93,RTX 2080 SUPER 移动版)裸跑的目标版:直接换 EFI\BOOT\BOOTX64.EFI,无需任何 marker。预期日志序:STAGING-FALLBACK: booter 0x89 ... 门控行 → FLR done (cold engine, WPR2 wiped) → FWSEC re-ignited, WPR2 re-verified → LOW 带重建 → 开火 → SELECTOR RESULT / HANDOFF-OUT ... tu10x:unlock=true。
以下为自 v2.1.11(上一稳定版)以来的全部变化(rc1 已含 1–8,rc2 新增 9–10;跳过 rc1 的用户请整表读)。
解锁行为变化
1. STAGING-FALLBACK 自动回退(rc1 引入):Turing 载荷轮 booter mailbox0=0x89 且宿主 meta magic 完好 = ≥4GiB staging 读未送达(#74/#156/#191 类平台)→ 自动单次 LOW 带重试,不再需要手工 \tx-low-staging.txt。0x91 镜像类(sub-4GB 读锁,#16/#57)与宿主页损坏(buffer-side)按设计不进门;单发无循环,两带都死走原失败路径 + VT-d 提示。
2. rc2 修正重试载具:1e93 现场证伪了 rc1 的"engine reset 恢复 fire 面"假设——完整跑过一轮 booter(halt+sanitize)后 FBIF 写面死掉(falcondmaunblock 写 0x115/0x0 读回 0x114/0x80,EIO 确定性)。rc2 重试改为:function-level FLR(整 function 冷态,FBIF 一并复活)→ FWSEC 重新点火(FLR 擦 WPR2 须重建)→ LOW 带重建 → 开火。门开火点零解锁、零 payload 写,FLR 无状态可毁。0x89 机上重试段代价 ~2–3s(正常机零成本:门条件未变,健康 HIGH 机零误触,台架证据)。sim 新故障 fbifwritedeadafterrun + t40 场景建模该平台(从 fallback 删 FLR 步即以现场 EIO 原形失败),t34–t40 全绿。
3. 三代交棒默认门开(FEAT PLM 不再交棒前重锁):TU106 bench 端到端坐实——卡自身 GOP 重建与 Windows 驱动 init 都不重关门,系统内 PLM 保持 0xffffffff,写入面开放(ssdiag --reapply 的 Turing 只读限制具备解除条件)。\tx-relock.txt 恢复旧重锁行为(Turing 臂回锁值 0xffffffef/0xffffff8f 随板方差;Ampere/Ada 臂未 bench)。
速度/静默旋钮(全部 opt-in marker,缺省零行为差异)
4. \tx-ultrafast.txt 速度档(已验证机器用):跳过全部纯诊断探测(M1 冷态采样/GOP 探测/AUTH-SCAN/ROOT-DIR 转储/opener 拓扑/载荷后 FEAT sweep/HANDOFF-PCI/HANDOFF-VGA-FINAL/POST-GOP 探针/CHAIN-MEMMAP),折叠 quiet + 仅事务锚点 flush(~80→~15 次/轮);复位金丝雀保留(HANDOFF-OUT 带真值);CHAIN reconnect pass 一并跳过。裁决读回一个不少:selector 写读回、WPR2/POST-BOOT 判定、booter 载入审计照跑。
5. \tx-no-log.txt / \tx-no-console.txt:分别不开 \loader.log / 不回显;同置 = 完全静默跑(no-console 确认提示不可见,必须配 armed,键盘语义不变)。
6. armed 倒计时接受 0:\iknowwhatiam_doing.txt 首行 0 = 不等待,只读预备完成立即 PROCEED(此前钳到 1 秒)。
7. armed/chain 路径两条固定 stall 删除(~550 ms)。
8. OS-loader 预载改 opt-in(#198 交棒后挂死嫌疑收敛):\tx-preload.txt 启用预载;缺省不预载,BootOrder 引导链的索引解析与跨盘防护保留。
9. 旧 banner 串修正('unloader'/'M0 skeleton' 残留)+ banner cstr16 导入按 feature 门控。
实测与工具
- 速度阶梯收口(Z170/TU106 台架,内置 SATA ESP,armed+quiet+ultrafast+no-console):v2.1.9 quiet 代 17.2 s → ~1.9 s 全程(事务段 detach→FINAL ≈1.2 s);逐级证据在 docs/evidence/tu106-1f07/。
- targeted PROM 读实验两提交已 revert(净零代码变化)。
- tools/bench_v2.py:社区标准四精度 GEMM 测量工具 vendoring。
- v2.1.11 的 zip 已原地重打包移除内嵌 ssdiag(杀软误报政策 #190/D#189——ssdiag 只以独立附件分发)。
- ssdiag 0.7.2 独立附件(本 release 新增,SHA256SUMS 一并覆盖):修复 SR-IOV VF guest(vfio-pci,geforce-sriov 类栈)下的静默崩溃——VF 扩展配置空间读全 FF,扩展能力链走到 0xFFF 处的 u32 读跨出单页 ECAM 映射,进程在 ranges 之后无提示退出(67bb859)。现在读全部钳在单页内、全 FF 有专门报告行,标准链 link caps 照常打印。顺带正面数据点:pmxdrv 在 Secure Boot ON 的 guest 里装载成功。判读边界:VF guest 里的 UNLOCKED/NOTUNLOCKED 不构成宿主卡解锁证据**(VF 寄存器窗是 per-VF 过滤视图,FECS/PLM 属 PF 裸 PRI 状态);VFIO 整卡直通(#135)仍是唯一有效 VM 数据点。
资产与校验
| 资产 | 大小 | sha256 |
| --- | --- | --- |
| BOOTX64.EFI.2.1.12-rc2 | 462,336 B | 36151d3203310b7df949b284b434b5cd174cdf2eb6af1f0be27569f3732be8af |
| BOOTX64.EFI.readonly.2.1.12-rc2 | 189,440 B | 7dcae3389b1241bf307a225e063c2c95c0c9f23f72bc3f98d19843a16005dce9 |
| ssdiag-x86_64-windows-0.7.2.exe | 371,200 B | a4c6984fde2bdee1f3a791fe8d40fed669c0fe8f9979e090448ccf5a2834ef5c |
TX 构建经 check_efi.py 终验(TX banner YES / readonly YES 各自匹配)。Release: tag v2.1.12-rc2 = main 的版本 bump 提交。安全宪法副本与回滚手册见 README;解锁为易失覆盖(断电即失、每次引导重放),黑屏/STOP 即断电走正常引导。
---
v2.1.12-rc2 (pre-release) — STAGING-FALLBACK retry vehicle fixed (FLR + FWSEC re-ignition)
This is a pre-release; Latest remains v2.1.11. This is the target build for #156 (1e93, RTX 2080 SUPER Mobile): swap EFI\BOOT\BOOTX64.EFI and run bare — no marker needed. Expected log order: the STAGING-FALLBACK: booter 0x89 ... gate line → FLR done (cold engine, WPR2 wiped) → FWSEC re-ignited, WPR2 re-verified → LOW-band re-stage → fire → SELECTOR RESULT / HANDOFF-OUT ... tu10x:unlock=true.
Everything below is the complete delta since v2.1.11 (the last stable) (1–8 shipped in rc1, 9–10 are new in rc2; if you skipped rc1, read the whole list).
Unlock behavior changes
1. STAGING-FALLBACK auto-retry (introduced in rc1): a Turing payload run dying on booter mailbox0=0x89 with the host meta magic intact means the ≥4GiB staging read did not deliver (#74/#156/#191-class platforms) → one automatic LOW-band retry, no manual \tx-low-staging.txt needed. The 0x91 mirror class (sub-4GB read lock, #16/#57) and a corrupt host page (buffer-side) never enter the gate by design; one shot, no loops, and a both-bands-dead machine falls to the canonical failure path with the VT-d hint.
2. rc2 fixes the retry vehicle: 1e93 field-falsified rc1's "engine reset restores the fire surface" assumption — after one completed booter run (halt + sanitize) the FBIF write surface is dead (falcondmaunblock writes 0x115/0x0, reads back 0x114/0x80, deterministic EIO). The rc2 retry is: function-level FLR (the whole function cold, FBIF included) → FWSEC re-ignition (the FLR wipes WPR2) → LOW-band re-stage → fire. At the gate nothing is unlocked and zero payload writes happened, so the FLR has no state to destroy. On a 0x89 machine the retry costs ~2–3 s (zero cost on healthy machines: the gate is unchanged, HIGH-reachable machines never enter it — bench evidence). New sim fault fbifwritedeadafterrun + scenario t40 model the platform (removing the FLR step from the fallback fails t40 with the exact field EIO shape); t34–t40 all green.
3. All three families now hand off with the FEAT gate OPEN (no pre-handoff relock): a TU106 bench proved end-to-end that neither the card's own GOP rebuild nor the Windows driver init re-closes it — in-OS PLM stays 0xffffffff and the write surface stays open (the ssdiag Turing read-only --reapply restriction can now be lifted). \\tx-relock.txt restores the legacy relock (Turing-arm value 0xffffffef/0xffffff8f varies by board; the Ampere/Ada arm has never been benched).
Speed / silence knobs (all opt-in markers; zero behavior change when absent)
4. \\tx-ultrafast.txt speed profile (for machines with several verified rounds): skips every purely diagnostic probe (M1 cold sample/GOP probe/AUTH-SCAN/ROOT-DIR dump/opener topology/post-payload FEAT sweep/HANDOFF-PCI/HANDOFF-VGA-FINAL/POST-GOP probe/CHAIN-MEMMAP), folds in quiet + transaction-anchor-only flush (~80 → ~15 flushes/run); the reset canary stays (HANDOFF-OUT carries real values); the CHAIN reconnect pass is skipped too. None of the verdict readbacks are skipped: selector write/readback, WPR2/POST-BOOT verdicts and the booter load audit all still run.
5. \\tx-no-log.txt / \\tx-no-console.txt: never open \\loader.log / no console mirror at all; together = a fully silent run (no-console hides the confirm prompt — combine with armed; keyboard semantics unchanged).
6. Armed countdown accepts 0: first line 0 in \iknowwhatiam_doing.txt = no wait, PROCEED as soon as the read-only prep finishes (previously clamped to
v2.1.12-rc1 — STAGING-FALLBACK: automatic LOW-band retry for unreachable ≥4GiB staging (#74/#156/#191)预发布2026-11-08
nv-uefi-unlocker v2.1.12-rc1
STAGING-FALLBACK:≥4GiB staging 读不可达时的一次自动低位重试 / One automatic LOW-band retry when ≥4GiB staging reads go undelivered
> Pre-release(候选版)。默认行为在健康机器上与 v2.1.11 逐字节一致;台架双臂已于 2026-11-09 跑毕全绿(见下),现欠 0x89 实机的现场回收。
⚡ 性能:从 loader 起跑到交棒 Windows 引导链 ≈2.8 秒 / Speed: ~2.8 s from loader start to the Windows boot chain
维护者台架实测(TU106 + 引导盘为内置 SATA Samsung 850 EVO 1TB,速度栈 = \iknowwhatiamdoing.txt 首行写 0 + \tx-quiet.txt + \tx-ultrafast.txt + \tx-no-console.txt):log 113ms → CHAIN: LoadImage 2835ms,全程 ≈2.835 秒**。
该数值与 GPU 数量及整机硬件配置直接相关,本节实测机的配置:主板 = ASUS Z170-A,CPU = Intel 10980HK ES(QTJ1 魔改——笔记本核心上 1151 针脚转接座)@ 4.7GHz,内存 = DDR4-2720 MT/s @ 15-18-18-34,GPU = 单张 RTX 2070(TU106)。多卡机器逐卡串行跑完整事务会按卡数累加;更慢的存储/内存/CPU 会让资产读取与 FWSEC/BOOTER 阶段相应拉长——2.8 秒是"单卡 + 本节硬件 + 速度栈全开"的读数,不是全机器承诺值。
同机演进阶梯(每级参照系都给出):
| 阶段 | 实测 | 参照 |
| --- | --- | --- |
| v2.1.8 代(诊断全开、全量回显) | 22.1 s | issue #38 基线(同一台架) |
| v2.1.9(quiet + armed 5s) | 17.2 s | issue #38 实测(同一台架) |
| rc1(ultrafast + no-console + armed 1s) | 3.2 s | 本版首日实测(log 113→3205ms) |
| rc1(armed 首行 0) | 2.835 s | 本节(log 113→2835ms,0s countdown elapsed - PROCEED @1187ms) |
快在哪里、什么没有被省掉:\tx-ultrafast.txt 跳过的全部是纯诊断探测(M1 冷态采样、GOP 探测、AUTH-SCAN、ROOT-DIR 转储、FEAT sweep、POST-GOP 探针、CHAIN-MEMMAP、金丝雀);\tx-no-console.txt 把慢速 GOP 文本回显的成本归零;flush 稀化到事务锚点。资产照旧全量读入内存,裁决链读回一个不少——selector 写读回、WPR2/POST-BOOT 判定、booter 载入审计、HANDOFF-OUT 全在日志里:速度不是用验证换的。armed 0 = 只读预备一完成立即 PROCEED,中止窗 = 预备窗本身(预备期间按键仍中止)。
背景:staging 带要求按机器双向分化 / The staging-band requirement is machine-dependent in both directions
Turing 20 系上,booter 对 WprMeta/signature/radix staging 的 DMA 读取可达性存在两类方向相反的平台实证:
| 类 | 实证机器 | HIGH(>4GiB)staging | LOW(<4GiB)staging |
| --- | --- | --- | --- |
| ≥4GiB 不可达 | #74 1e04 2080 Ti 22G(TB4 eGPU)、#156 1e93 2080S-M(DGFF,Area-51m R2,首个非 TB)、#191 1ed1 2070S Max-Q(Legion 7 内屏) | booter 0x89 | ✅ |
| sub-4GB 读锁(反方向) | #16 1f14 2070(Clevo 准系统) | ✅ | booter 0x91 |
| 两头堵 | #57 1e02 TITAN RTX | 0x11 | 0x91 |
| 多数机器(含维护者台架 TU106) | — | ✅ | ✅ |
0x89 的判据链(GA 明文码表 0x2784=二次魔数比对失败 + 宿主侧 magic 完好):写侧落盘、读侧未送达——这正是"≥4GiB 不可达"类的唯一定义性签名,也是本版自动回退的门控条件。
新行为 / New behavior
默认(HIGH staging,driver-exact,不变)下,若 booter halt 时 mailbox0=0x89 且宿主读回 meta magic 完好(0xdc3aae21371a60b3):
1. 释放死掉的高位缓冲,整组 staging(meta/signature/radix+RM image/boot image)经 \tx-low-staging.txt 同一条分配路径重落低位(booter 镜像本身走 portal 不走 DMA,原地不动);
2. 单次重点火,无 FLR——此刻 PLM 未开、零解锁写入,无状态需要保护,fire 自带 B-reset;
3. 成功 → 日志 STAGING-FALLBACK: LOW-band retry verified - staging rebound to the LOW buffers (staging=low(fallback)),事务照常走完;
4. 回退也失败 → 走既有失败路径并打印提示(VT-d 开则两个带都不可达(#74 定案)——查 BIOS VT-d 与 DMA-PATH BME 行)。
门控拒绝(不触发回退,直接走原路径):0x91/0x8a/0x0B/0x11/0x31 等非 0x89 状态、宿主 magic 损坏(缓冲侧 bug)、低位已强制(\tx-low-staging.txt 在位)、\tx-force-continue.txt 实验位。每次引导至多一次重试,无循环。
旋钮不变:\tx-low-staging.txt 仍可直接强制低位(跳过整个回退路径);\tx-fwsec-high-staging.txt 照旧。GA10x/Ada 不经此路径(其低位默认本就经田野验证)。
新控制门随本版进入发布线 / New knobs shipping in this build
这三个旋钮此前只在 main/本地构建存在,本版起随发布线交付(完整行为表见 FAQ 标记表):
| 标记 | 行为 |
| --- | --- |
| \tx-ultrafast.txt | 速度档:跳过全部纯诊断探测(M1 冷态采样、GOP 探测、AUTH-SCAN、ROOT-DIR 转储、载荷后 FEAT sweep、POST-GOP 探针、CHAIN-MEMMAP、复位金丝雀),折叠 quiet + 仅事务锚点 flush(~80 次/轮 → ~15 次/轮)。裁决链读回一个不少:selector 写读回、WPR2/POST-BOOT 判定、booter 载入审计照跑——读回是唯一成功判据 |
| \tx-no-console.txt | 完全不回显(慢速 GOP 文本输出的回显成本归零)。确认提示不可见——必须与 armed 同用(armed 倒计时内任意键仍中止);与 \tx-no-log.txt 同置 = 全程零输出,挂死时屏与盘两处都无证据,只建议已多轮验证的机器用 |
| \tx-no-log.txt | 不开 \loader.log(全程仅 console 输出)。挂死即零取证,同样只与 armed + 已验证机器同用 |
实测速度 / Measured speed
首日 3.2 s(armed 1s)与 2.835 s(armed 0)两轮的完整日志都在证据目录 docs/evidence/tu106-1f07/rc1-staging-2026-11-09/(commit 记录见本仓库 main)——演进阶梯与口径见文首"⚡ 性能"节,此处不再重复。
交棒行为变化 / Chain behavior changes
- OS loader 不再预载内存,但高级启动项索引解析原样保留:v2.1.9 起曾把 bootmgfw 提前 LoadImage(TOTORONG 工艺 P2-7),#198 九格矩阵中"预载 + 脏卡"与每轮交棒后硬挂死逐一对应,故本版起默认回到事务后链载,\tx-preload.txt 变 opt-in(复现旧行为用)。预载原有的鲁棒性动机(FLR 后分配面退化时链载失败)由固件 BootOrder 索引解析链覆盖(Pass 0.5,#128/#142 双 ESP 修复;跨盘残留判例 #204 的防护不变)——链载目标解析的鲁棒性没有随预载一起退场。
- armed 秒数 0 合法:\iknowwhatiam_doing.txt 首行写 0 = 不等待——只读预备(枚举/选卡/资产读入)一完成立即 PROCEED,中止窗 = 预备窗本身(预备期间按键仍中止);此前钳到 1 秒。
本版同时携带:Turing 交棒默认门开(不再 relock)/ Also in this build: Turing hands off with the FEAT gate OPEN
v2.1.11 及以前,Turing 交棒前默认把 FEAT PLM relock(日志 TU10X PLM relock: plm=0xffffffef,bits 5:6 sticky-open)。本版起三代一致:默认门开交棒——SS 覆盖值本就是载荷本体且为 AON(FLR 都打不回),门开把系统内写入面开放出来(ssdiag --reapply 的 Turing 限制具备解除条件),TU106 台架端到端坐实门开无害:系统内 ssdiag 实测 PLM 回读 0xffffffff(OPEN)、SS=88888888、READOUT=0x00000000 full speed(见 docs/evidence/tu106-1f07/keep-open-2026-11-08/)。想恢复旧的交棒锁门形态:\tx-relock.txt(三代通用退出阀)。判读注意:HANDOFF-OUT 尾部重读的四个 FEAT 字段(ss0/ss1/readout1/featplm)在 native-GOP console 重建成功绑定后会读到 0xbadf5040——这是卡自家 GOP 驱动绑定期的读窗 artifact,门开与 relock 两臂都有(权威判读=同一行的 tu10x: 元组与系统内 ssdiag**,#125 教训),不代表门被重关。
为什么默认仍是 HIGH / Why the default stays HIGH
真驱动把 meta stage 在 >4G(live dump 0x110BB0000)——HIGH 是 driver-exact 形态,保护的是 #16 那一类 sub-4GB 读锁机器;翻 LOW 等于拿一类换另一类。自动回退让"高位不可达"类不再需要手工定位,同时 STAGING-FALLBACK 行保持该类分布数据可见(不被静默修复吃掉)。
判读速查 / Log reading
- STAGING-FALLBACK: ... one retry in the LOW band — 回退已触发(你的平台属于上表第一类);
- staging=low(fallback) — 回退成功,解锁正常交付;
- LOW-band retry failed as well — 两个带都不可达:优先查 VT-d(开则全死,#74 四格矩阵),其次桥 BME;
- 未出现任何 STAGING-FALLBACK 行 = 默认路径一次通过(多数机器)。
验证状态 / Verification status
- sim:t34–t39 六场景新增(回退恰一次恢复 / 两头堵收敛 / 0x91 镜像类不触发 / magic 损坏不触发 / LOW 强制负控 / 健康多数不进门),全套 172 passed;
- 构建:tx 与 readonly 双变体 check_efi 终验通过(tx 460288B / readonly 188416B);
- 真机(2026-11-09 已跑,双臂全绿):TU106 台架双臂收官——HIGH 负控:meta 落 0x10001f000(与 #156 毒机完全同段)mailbox0=0x0 全绿,全程零 STAGING-FALLBACK 行(健康机门零误触);LOW 臂(\tx-low-staging.txt):全 LOW 分配、全绿、零回退行,且系统内 ssdiag 实测 PLM=0xffffffff(OPEN)+ SS0=0x88888888 + READOUT=0x00000000(full speed)——rc1 门开交棒在 OS 侧活体确认。证据入档 docs/evidence/tu106-1f07/rc1-staging-2026-11-09/(commit 3b6c2de)。台架无法行使的只剩一条路径:真 0x89 机器上回退真的开火——手上是 0x89 机器的报单者:此构建值得优先试(#156/#191/#74 任一 issue 附日志即可)。
下载 / Downloads
- BOOTX64.EFI.2.1.12-rc1 — tx 构建(含解锁事务;SHA-256 见 SHA256SUMS.2.1.12-rc1)
- BOOTX64.EFI.readonly.2.1.12-rc1 — 只读构建(无 GPU 写路径,日志/诊断协议同前)
- 制盘步骤不变(make-usbroot + 资产生成照 v2.1.11 文档);只换 EFI\BOOT\BOOTX64.EFI 即可从 v2.1.11 升级。
- 资产换件记录(2026-11-09 第二版):本页 EFI 于 rc1 发布数小时后原位重传,新增 armed 0 秒支持(STAGING-FALLBACK 逻辑零改动,仍为 e2217b2 同源)。新 tx sha fffde6bc… / readonly d4c2c097…,以本页 SHA256SUMS.2.1.12-rc1 为准;此前引用 bece7118…/1ee6246a… 的旧 hash 作废。
Full changelog: https://github.com/physicsdolphin/nv-uefi-unlocker/compare/v2.1.11...v2.1.12-rc1
v2.1.11 - local-driver GSP default (badge mode) + measured 525-620 GSP compatibility2026-11-08
nv-uefi-unlocker v2.1.11
v2.1.11 — 制盘下载与 GSP 来源的结构性升级(本地优先)+ TU106 台架五版本 GSP 谱系实测
主要变化 / What's new
1. 本地驱动 GSP 成为默认资产来源(badge 模式)/ Local driver GSP is now the default asset source ("badge mode")
多数制盘机都装着 NVIDIA 驱动——驱动仓库里本来就带着分代 GSP 固件(Windows C:\Windows\System32\DriverStore\FileRepository\nvinfamd64\gspga10x.bin + gsptu10x.bin;Linux /usr/lib/firmware/nvidia/<版本>/gsp.bin,NixOS 等另探 /run/current-system/firmware/)。v2.1.11 起 make-usbroot 默认优先取本地驱动的 GSP,找不到才回退网络下载:
- Windows:扫描 DriverStore,从 .inf 解析 DriverVer,多仓库取最新;
- Linux:/usr/lib|/lib|/run/current-system firmware 下 sort -V 取最高版本;
- 本地路径 = badge 模式:版本钉定豁免(实测依据见下),fwimage+fwsignature 由同一容器天然成对;
- pre-52x 老驱动(单一多代 gsp.bin,无 .fwimage 段)会被指名拒绝并回退下载链;
- 显式参数:-UtilsSource arch|run|cachyos|local(bash --utils-source)可钉定任一通道,local = 只用本地驱动 GSP;下载通道的哈希闸语义不变。
效果:装了 N 卡驱动的机器上,315 MB 的 nvidia-utils 下载默认消失,只剩 26 MB kernel-source(booter/BL/scrubber 的 BINDATA 内嵌在 nvlddmkm.sys 里,不是可拷贝文件,仍走钉定下载)。实测兼容的旧 assets 二进制同样可用(gsp-sections 早已存在,无新依赖)。
Most stick-building machines already have the NVIDIA driver installed, and the driver store ships the split GSP firmware ELFs. make-usbroot now tries the LOCAL driver's GSP first and falls back to the pinned downloads only when no local store is found: auto = local DriverStore (badge mode) -> Arch archive -> first-party NVIDIA .run -> CachyOS cache; pin one route with -UtilsSource arch|run|cachyos|local. Local is badge mode (version pin waived; fwimage+fwsignature paired by construction; pre-52x single-ELF stores are refused with a named reason and fall through). Net effect: no 315 MB download on driver-equipped hosts - only the 26 MB kernel source still downloads (booter/BL/scrubber BINDATA is embedded in nvlddmkm.sys, not a copyable file).
2. GSP 在解锁链里是什么 / What the GSP actually is in the unlock chain
GSP-RM 镜像在本管线中的全部角色:砸门的工牌。payload 模式下,WprMeta 的 signature DMA 指向 V67 载荷,booter 的签名拷贝越界即触发 ROP——对 GSP 镜像自身的任何校验都在那之后(门已开)。事务对镜像的全部依赖面是三个只吃"尺寸"的纯函数(layout / radix / WprMeta image_bytes)。真正的版本咬合三件套——booter、GspRmBoot BL、FWSEC——全不来自 nvidia-utils 包。机制详见 docs/gsp-compatibility.md。
In the unlock chain the GSP image is just the badge that smashes the FEAT gate open: in payload mode the WprMeta signature DMA aims at the V67 blob, so the booter's own verification of the image never runs before the ROP fires. The transaction's entire dependency surface on the image is three pure functions fed the SIZE. The parts that genuinely bind to versions (booter, GspRmBoot BL, FWSEC) do not come from the nvidia-utils package at all.
3. TU106 实测:525–620 GSP 全谱系通过 / Measured: the whole 525-620 GSP spectrum unlocks (TU106)
Z170-A/TU106 台架实测,五个同格式 GSP 全部全链解锁(selector 0x88888888/0x8 落盘、READOUT_1=0x0 全速、POST-GOP unlock intact):
| GSP | fwimage 大小 | loader |
|---|---|---|
| 525.60.11(最老同格式) | 37,068,768 B | main v2.1.11 |
| 560.35.03 | 26,250,912 B | v2.1.10 |
| 610.43.03(钉定 stock) | 29,339,080 B | v2.1.10 |
| 615.78.08 | 22,887,880 B | v2.1.10 |
| 621.01(Windows 驱动仓库,即交棒对象自带固件) | 16,674,264 B | main v2.1.11 |
尺寸跨度 2.2×,版本、尺寸、出处全部非因子。作用域按世代限定:以上为图灵谱系 gsptu10x;各代 GSP 只用于本代卡(40 系/AD10x 走 gspga10x.bin 的 fwsignature_ad10x 段,下限=40 系起家最早 GSP)。唯二硬边界:容器格式代际(510–520 是单一多代 ELF,无 .fwimage 段,结构上不可打包);fwimage↔fwsignature 必须同容器成对。
On a Z170-A/TU106 bench, five same-format GSP firmwares across the whole 525-620 spectrum (sizes 16.7-37.1 MB, including the handoff-target's own driver-store GSP) unlock end-to-end with selector read-backs and the POST-GOP verification intact. Scoped per generation: this is the Turing lineage (gsp_tu10x); each generation uses its own GSP. Hard limits unchanged: container generation (510-520 predate the split format) and pairing.
4. 两个 main 回归修复(本轮台架暴露并验证)/ Two main regressions found and fixed by this bench round
- 日志管线:diag3-5 时代的 marker Close/reopen 在部分固件(Z170-A/#198 族)上被拒绝 → 首条 durable 行就永久丢失文件句柄 → 整场 0 字节盲跑。修复=移植 diag7 的 log-commit 阶梯:默认 flush-only(v2.1.9 语义、单长句柄),reopen 被拒时自动降级;tx-log-commit-lines/tx-log-commit 显式 opt-in 旧诊断行为。
- TU10x verdict:#74-round-33 的"按 payload 写 SS 判"锁死了另一种同样合法的成功形状(payload 只开 PLM、loader 主机写 SS——v2.1.10 全部 TU106 成功轮的分工)。修复=双形状判定:PLM 开 + SS 工厂值 → 交给主机写路径,读回落盘才是判决(台架 621.01/525 双连验证解锁)。
- Log pipeline: the diag3-5 marker Close/reopen is refused by some firmware (Z170-A, the #198 family), silently dropping the file handle on the first durable line - the run went blind at 0 bytes. Fixed by porting the diag7 log-commit ladder: flush-only is the default again (v2.1.9 semantics), with automatic degrade when the reopen chain refuses; the diagnostic behavior is opt-in via tx-log-commit-lines / tx-log-commit.
- TU10x verdict: the #74-round-33 change accidentally locked out the other proven success shape (payload opens ONLY the PLM; the loader writes the selectors as host). Fixed with a two-shape verdict; the host write and its read-back decide - bench-proven unlocking with both 621.01 and 525 GSPs.
5. 交棒清理默认化确认 / Handoff cleanup confirmed default
自 v2.1.10 系起,交棒前资产池释放即为默认行为(本轮两份台架 log 均见 TX-ASSET-CACHE: cleared N entries, ~17-37 MB handed back to the firmware allocator,交棒低位内存最大块 301 MiB)——#146 的 winload 0xc000009a(资产池劈碎唯一低块)不再依赖用户手动释放。\tx-keep-assets.txt 仍为诊断退出阀。
Since the v2.1.10 line, releasing the asset pool before the handoff is the DEFAULT (both bench logs this round show TX-ASSET-CACHE: cleared ... handed back to the firmware allocator; largest low block at handoff 301 MiB). The #146 winload 0xc000009a class no longer depends on manual cleanup; \tx-keep-assets.txt remains the diagnostic opt-out.
6. 下载层韧性 / Download resilience (carried from main, now shipped)
断点续传(curl.exe 优先,失败原因实报)、两种失败判语分家("没有源给出了钉定字节" vs "源给出了错误字节")、全败时列出全部源+URL+钉定哈希+离线投放指引。本轮全程实测:Arch 归档整包、本地 DriverStore、断点续传恢复、三种失败判语。
Range resume (curl.exe preferred, every failure reason logged), two distinct failure verdicts, and an all-sources-failed report listing every source, URL and pinned hash. All paths bench-exercised this round.
Credit
- #180(j0p3s):下载韧性的诊断起点("China-only 框架太窄"的判断正确且被 #205 数日内证实)。
- #212(ray7978-lang):615.78.08-2 pkgrel 观察直接定位了 CachyOS 滚动窗机制。
- #205(epokuser):第一份完整的"传输静默死亡"现场记录(exit 35 指纹)。
- 台架回归的定位依赖 #170/#198 的簇雕刻与 0 字节 log 形态学(fail-flag 纪律)。
SHA-256 / Checksums
见 SHA256SUMS.txt(zip 外层与 zip 内各一份,内容一致)。BOOTX64.EFI = 448,512 B,TX 构建(check_efi.py 验证通过)。Release: tag v2.1.11 = main 9b02f2e(后续 docs 修正 4ad6cfc 不影响构建产物)。
资产换件记录(2026-11-09)
nv-uefi-unlocker-v2.1.11.zip 原地重打包:移除 zip 内嵌的 ssdiag-x8664-windows-0.7.1.exe(杀软误报政策 #190/D#189——ssdiag 只以独立附件分发,永不入 zip;此包当年漏执行该政策)。其余条目逐字节原样(BOOTX64.EFI/assets/make-usbroot/enable-turing-unlock/LICENSE/README 未动),内层 SHA256SUMS.txt 同步去掉 ssdiag 行并与其余条目逐一复验。新 zip sha256 dfc5644b7084d1ee804e0b25f43ddca92018cefcc1517dd5d4994d2481a92062;独立附件 ssdiag-x8664-windows-0.7.1.exe 与其余资产 hash 不变,外层 SHA256SUMS.txt 无 zip 行故不需改动。
Asset swap record (2026-11-09)
nv-uefi-unlocker-v2.1.11.zip repacked in place: the embedded ssdiag-x8664-windows-0.7.1.exe is removed (AV false-positive policy #190/D#189 - ssdiag ships as a standalone attachment only, never inside the zip; this package missed that policy). Every other entry is byte-identical (BOOTX64.EFI/assets/make-usbroot/enable-turing-unlock/LICENSE/README untouched); the inner SHA256SUMS.txt drops the ssdiag line and was re-verified against every remaining entry. New zip sha256 dfc5644b7084d1ee804e0b25f43ddca92018cefcc1517dd5d4994d2481a92062; the standalone ssdiag-x8664-windows-0.7.1.exe and all other asset hashes are unchanged, and the outer SHA256SUMS.txt needs no edit (it carries no zip line).
v2.1.11-diag5 — FWSEC ignition line-resolved (#170) [renamed from 2.1.14-diag5]预发布2026-11-06
Name note
This build was first published as 2.1.14-diag5; the diag line is now unified on 2.1.11-diagN (the diag number carries the increment, the base tracks the code series). Same code plus the two log-only fixes merged since (#185 backslash-usbroot fallback scoped to asset files, #186/#74 eGPU recovery note in the failure unwind); new hashes above.
Why diag3 and diag4 were withdrawn
2.1.12-diag3 and 2.1.13-diag4 have been removed from the releases page (releases and tags deleted, 2026-11-06) so nobody spends cold boots on their judgement tables. Reason: both were built around a premise that the FAT cluster-chain recovery has since overturned - the recovered v2.1.10 logs prove the transaction runs through the detach (5.6 s) into AUTH-SCAN, the scrubber and the layout, so the mute-split ("does the hang survive the console mute at DisconnectController?") was splitting a segment nothing dies in, and diag4's "first-line window" reading of a 0-byte log does not survive the Flush-noop reality either (a 0-byte file means only "no Close yet", not "died early").
Everything in them that remains valid - durable marker lines, per-line commits from line 1, the DIAG-CONSOLE/DIAG-BOOTVOL identity dumps, the early connect pass, the mute across the detach - is carried by this build, with the judgement re-aimed at the FWSEC ignition stretch. The four recommending comments on #170/#122 are patched in place with withdrawal banners; the protocol doc (docs/diag-2.1.12-diag3-protocol.md) marks its diag3/diag4 sections as historical.
What this build is
2.1.11-diag5 exists because of outstanding forensics from the #170 reporter: two complete v2.1.10 logs recovered from the stick's FAT cluster chain. They show the transaction completed the detach at 5.6 s, ran through AUTH-SCAN, the SEC2 scrubber (ETimeout, tolerated) and the WPR2 layout, and the evidence ends at TX: layout ok (~10.9 s).
Two corrections follow, and both are baked into this build:
1. The "hang at DETACH" premise is withdrawn. The screen's last painted line was the console-teardown freeze, not the execution endpoint. The mute split in diag3/diag4 was aimed at the wrong segment on this board.
2. "Stopped at layout ok" is itself not established. TX: layout ok and the next log call are adjacent statements - nothing between them can hang. Under the v2.1.10 evidence pipeline (where flush never commits the dirent on this board), the absence of later lines is weak evidence. Execution very likely continued; the record just could not follow it.
What changed
The FWSEC ignition stretch is now line-resolved - every step emits an FWSEC:-prefixed commit line:
``
FWSEC: reading PROM at 0x300000
FWSEC: PROM read 25% / 50% / 75% / 100% <- the 1 MiB bulk read, now 16 chunks (behaviorally identical)
FWSEC: PROM bytes=... / imem=... / stored=...
FWSEC: image staged (... frts patched)
FWSEC: fuse raw=... version=...
FWSEC: GSP falcon reset begin
FWSEC: GSP reset+unblock ok
FWSEC: IMEM portal begin (N blocks)
FWSEC: portal IMEM done
FWSEC: DMEM portal done (...)
FWSEC: PKC params + BOOTVEC writes begin
FWSEC: HALT wait begin (5s cap)
FWSEC: portal forensic CPUCTL=... MAILBOX0=... EXCI=...
FWSEC: result scratch=... / WPR2 low=... high=...
`
TX: and layout join the commit-marker set, so every segment boundary lands on the stick. No new hardware operations - same registers, same order; one bulk read split into 16 identical reads. Also includes everything from diag4 (per-line commits from line 1, console mute, identity dumps) and is rebased on main including the #109 capture-wording fix.
The screen will freeze - by design
Both reporting boards (#122 and #170) lose their console inside the GOP DisconnectController call: the submitted screen stops at DETACH: stage GOP handle while the log keeps going for another 1-2 lines (32 recovered #122 sessions, both #170 sessions). After that the console still returns from TextOut but paints nothing.
This build mutes the console mirror one line before that disconnect, so every run freezes on DETACH: console mirror MUTED until reconnect, green or dead. A frozen screen is no longer a hang signal - judge by the log only, and ignore the screen until the RECONNECT/CONSOLE: mirror re-enabled lines.
If you are the #170 reporter
Use BOOTX64.EFI.2.1.11-diag5 (tx). One cold boot first; read the last line against the table in docs/diag-2.1.12-diag3-protocol.md §5 (Chinese). Quick form:
| Last line | Meaning |
|---|---|
| stops at PROM read NN% | the BAR0 PROM window wedges mid-bulk-read - new failure shape |
| stops at GSP falcon reset begin / reset+unblock ok | reset propagation / FBIF unblock |
| stops at IMEM portal begin | the register-portal write loop |
| stops at HALT wait begin and nothing after | death inside the bounded wait's MMIO reads |
| portal forensic present | the wait ended; the three values decide what FWSEC itself reported |
If the dirent again fails to commit, your cluster-chain recovery is the fallback reader - it is now a proven method on this stick, and the build is designed so the recovered text lands on a step boundary.
With the localization in hand, A/B switches (one per cold boot) come after - not before. Note the GA10x FWSEC path has no staging marker (\tx-fwsec-vram-staging.txt is Turing-only), so that switch is not available here; and scrubber did not complete (ETimeout) is not a verdict for GA10x - the GA106 2504 full-chain success had the same shape, so no workaround is needed or exists for it.
Also requested with the next round
- the two recovered logs, the raw 32-byte dirent, and the 78-orphan-cluster list attached to the issue (you offered - please do; they are the evidence base for this re-orientation);
- the diag4-round screen photo: whether the banner appeared or the screen was black from the start. The diag4 round's committed-but-empty dirent is ambiguous between "the app died before the first line" and "the app never launched this boot" (a stale dirent from a previous round looks identical); the screen discriminates.
Assets
| Asset | SHA-256 |
|---|---|
| BOOTX64.EFI.2.1.11-diag5 (tx) | a829fd439667e41510fdcdcd8dca4cd109e87a1c5257bdc3fc28bd68495577ce |
| BOOTX64.EFI.readonly-diag.2.1.11-diag5 (control) | cd3f6ae4fcfb1801e0dfd2a9a7dc1da4789fd36efd8503d7cadb95aceb36c160 |
Replace only EFI\BOOT\BOOTX64.EFI. SHA-256 also in SHA256SUMS.2.1.11-diag5`. Rollback: put the previous binary back; nothing persistent is written.
v2.1.11-diag7 — auto-degrade when the reopen chain refuses (#198)预发布2026-11-06
What this build fixes
2.1.11-diag5 (the build originally tagged 2.1.14-diag5) loses the log entirely on the Y9000P (#198 secondary finding): loader.log stays 0 bytes on every run - tx and readonly alike - and a full-cluster carve finds no log content anywhere, even on the readonly run that completed successfully. The code path explains it: the diag3+ create-commit dance is CreateReadWrite → Close → Open(ReadWrite) → setposition(EOF), and on that firmware the reopen chain refuses cleanly** - the logger then degrades to console-only for the whole run, so nothing ever enters the file. The same Close/reopen machinery that wedges the Z97-K init window (#122, 4/4 no-signal) fails in a second, cleaner mode on this Lenovo. v2.1.10 and diag2 (no dance: one long-lived handle, flushes, a single Close at handoff) log fine on both machines - which is exactly why they were chosen as the reference.
2.1.11-diag7 keeps the ladder from diag6 and adds auto-degrade: if the reopen after the create-commit returns nothing, the logger recovers at init - nothing has been written yet - by taking a fresh handle and forcing commit level 0 (the v2.1.9 flush-only write path) for the rest of the run. Line 3 of every log names the state it ended up in:
``
BOOT: diag7 log-commit=N (0=v2.1.9 default, 1=+marker Close, 2=+create Close; auto-degrades to 0 if the reopen chain refuses)
`
Switches (unchanged from diag6)
| File on the stick root | Effect |
|---|---|
| (none) | default: write + Flush on marker/cadence lines, one long-lived handle, no Close (v2.1.9 semantics) |
| tx-log-commit-lines.txt | + Close/reopen on marker lines |
| tx-log-commit.txt | + the create-time Close (diag4/5 semantics, with auto-degrade) |
| tx-no-ident-dump.txt / tx-no-early-enum.txt | skip the identity dumps / the EARLY: connect pass |
What this means for the two open cases
- Y9000P (#198): default or tx-log-commit.txt both produce working logs now (the latter via the auto-degrade). Your runs should log like diag2 did - 47-56KB, committed at handoff.
- Z97-K (#122): the default (level 0) avoids the init-window wedge by construction; the ladder round stands as before (empty stick → tx-log-commit-lines.txt → tx-log-commit.txt, one cold boot per step).
The Close/reopen machinery now has two documented firmware failure modes; level 0 - the path every shipping release used - is the default everywhere and remains the reference.
Assets
| Asset | SHA-256 |
|---|---|
| BOOTX64.EFI.2.1.11-diag7 (tx) | f234056603341563a8eb2c19c49bb726c2bb89261212a627d5e6ff544ab78355 |
| BOOTX64.EFI.readonly-diag.2.1.11-diag7 (control) | 5c27e80b27390a5f20f61761f843e31af42433b6489a8f5340bb08af017ec133 |
Replace only EFI\BOOT\BOOTX64.EFI; hashes also in SHA256SUMS.2.1.11-diag7`. Rollback: put the previous binary back; nothing persistent is written. Transaction code is untouched - the only delta vs diag6 is the logging fallback.
v2.1.11-diag6 — log-commit ladder, v2.1.9 write path by default (#122)预发布2026-11-06
Why this build exists
The three diag5 cold boots on #122 died one layer above where diag5 was aimed: the log dirent was freshly created every round, yet zero bytes, zero FAT clusters, zero string hits and no display signal - the run ends inside the window that opens at the log's create and closes before the first line reaches the medium. The FWSEC judgement table (PROM/GSP/IMEM/HALT) is not reachable on that board until this window is crossed.
The suspect is the Close-based commit itself. diag3/4/5 always Close (once for the create, once per marker line); v2.1.9 and diag2 never Close at all, and both boards sedimented 157-239 lines per run on those builds. So this build makes the Close a stick switch and defaults it OFF - default behaviour is the v2.1.9/diag2 write path, byte for byte.
Switches
| File on the stick root | Effect |
|---|---|
| (none) | default: Write + Flush on marker/cadence lines, one long-lived handle, no Close ever (v2.1.9 path) |
| tx-log-commit-lines.txt | + Close/reopen on marker lines (diag3/4/5 commit half) |
| tx-log-commit.txt | + the create-time Close as well (full diag4/5 behaviour) |
| tx-no-ident-dump.txt | skip the DIAG-CONSOLE-*/DIAG-BOOTVOL dumps |
| tx-no-early-enum.txt | skip the pre-enumeration EARLY: connect pass |
Line 3 (BOOT: diag6 switches log-commit=N ...) records what was read, so every log names its own configuration. With commits off the dirent stays stale and the file reads 0 bytes on a crash - the FAT cluster carve is the reader, exactly as with v2.1.9.
The pre-log probes are themselves a probe
tx-log-commit is read before the log file exists. If a run dies inside those two directory probes there is no fresh dirent at all** - that distinguishes "died before the create" from "died after the create", which the media alone could not.
Run order (cold boot each, one round each, stop when the log comes back)
1. nothing on the stick - if logs sediment again, the Close is the regression; go to 4.
2. still nothing, but the dirent is not fresh → died in the pre-log probes.
3. still nothing with a fresh dirent → not the Close; add tx-no-ident-dump.txt, then tx-no-early-enum.txt, one per cold boot. Report which one, if any, restores the log.
4. if the log comes back: keep the level that works for the transaction runs, and add tx-log-commit-lines.txt on a later cold boot to see whether the per-line commit alone is survivable.
Report per round: whether \loader.log is fresh (dirent timestamp), its size, the first and last line (carve if 0 bytes), whether the BOOT: diag6 switches line is present, and the total elapsed between the first and last timestamp. The screen is expected to stay dark on this board - it is not a verdict.
Assets
| Asset | SHA-256 |
|---|---|
| BOOTX64.EFI.2.1.11-diag6 (tx) | 077b783299f995bba21d5396ae131c0812dc28b86a16abddcf692df7ce75017f |
| BOOTX64.EFI.readonly-diag.2.1.11-diag6 (control) | 14ed28440b40ecd7848915563f43b420041b824ad9a7daeb04cb09b664e067fd |
Replace only EFI\BOOT\BOOTX64.EFI; hashes also in SHA256SUMS.2.1.11-diag6. Rollback: put the previous binary back, nothing persistent is written. No transaction change - same registers, same order; only the logging commit path differs between levels.
v2.1.11-diag2 — #146 low-memory handoff fix is the default (#164 / #165 protocol unchanged)预发布2026-11-04
What this build is
2.1.11-diag2 = 2.1.11-diag1 + one behavioral change: the ~84 MB asset pool is handed back to the firmware allocator before the chain, on every path (issue #146). It fixes the project's first winload-stage 0xc000009a.
Root cause, measured with a three-run A/B on the reporting machine: ExitBootServices converts nothing to conventional, so our resident asset pool crosses into the Windows memory map as a LOADER_DATA descriptor in the middle of the only large conventional run below 4 GiB. winload takes its boot reservation out of a single descriptor and does not stitch - 2.4 GiB still free is not enough. Freeing restores one 2604 MiB run and the same machine, stick and BCD boot cleanly. Full write-up, the maps and the UEFI 2.10 / EDK2 basis: docs/handoff-lowmem.md.
What it does not change
Transaction chain, payload, FLR order, native-GOP rebuild, reset canary, POST-GOP / HANDOFF-VGA-FINAL probes, \tx-acpi-dump, POWER: readings: identical to diag1. If you are on #164 or #165, your protocol is unchanged - run exactly what you were given. The only new output is one number in the existing CHAIN-MEMMAP summary:
``
CHAIN-MEMMAP: ... largest block 2604 MiB (99 descriptors, 64 shown)
`
No GPU write was added; the release surface is the memory map only.
If you are the #146 reporter
This is the build to confirm on. Delete \tx-free-assets.txt from the stick root (now a no-op) and do one normal run. Expected, in this order:
`
TX-ASSET-CACHE: cleared 9 entries, 84388324 bytes handed back to the firmware allocator
CHAIN-MEMMAP: ... largest block 2604 MiB (...)
`
→ Windows boots without the Recovery screen. \tx-keep-assets.txt (empty file) restores the old behavior for an A/B; the log then says cache left resident across the chain. Send the whole loader.log either way.
Assets
- BOOTX64.EFI.2.1.11-diag2 - replace only EFI\BOOT\BOOTX64.EFI on the stick. Banner must read v2.1.11-diag2 (built <timestamp>) with TX BUILD.
- ssdiag.exe.0.6.8 - unchanged since diag1; in-OS read-back, run as administrator.
- ssdiag.exe.0.6.9 - adds the Blackwell (GB20x) read-only surface (issue #168): FEAT/fuse rows, gapped READOUT_1 decode, PL3-only verdict; --reapply refuses there. 0.6.8 stays above for the pinned #164/#165 protocol.
- ssdiag.exe.0.6.10 - adds the read-only per-GPC TPC floorsweep block for Ampere/Ada (issue #169): STATUS/CTRL pair, GPC disable/defect masks, defect-vs-candidate verdict.
SHA-256 in SHA256SUMS.2.1.11-diag2. Rollback: put the previous BOOTX64.EFI` back; nothing persistent is written by any build.
v2.1.11-diag1 (issue #164 / #165 diagnostics)预发布2026-11-04
Diagnostic build for #164 and #165. Not a release — the transaction is byte-for-byte the v2.1.10 chain; every addition is read-only sampling plus one two-word canary.
> Asset replaced a fifth time 2026-11-04 — adds the POWER: read-out (ssdiag 0.6.8). Current SHA-256 a85e282be230eb1413fe6718930a343103381b8858893d18f6b98777c738f107; take ssdiag 0.6.8 (attached) with it. Anything grabbed earlier is canary-less, discovery-buggy, carries the wrong canary offsets or lacks POWER: — re-download.
>
> Why: a TU106 run of the previous build came back fully green — canary intact on the loader side through HANDOFF-OUT and again in the OS (CANARY: intact + SS live), unlock live — which retires the reset hypotheses for that machine and leaves a plain write. POWER: closes the remaining physical axis: it reports the card's and its upstream root port's PMCSR / D-state, and the platform SSDT analysis (docs/evidence/tu106-1f07-acpi-dump/ANALYSIS-ssdt02-power.md) showed _OFF is PERST# + slot-rail-off, so a port in D3cold is the one power event that costs an unlock with no register write at all.
>
> Still in from the previous replacement, all found on a TU106 field run: canary at 0x1480/0x1484 (NVPBUSSWSCRATCH(i) = 0x1400 + i4, indices 0x20/0x21 — earlier builds read 0x140020/0x140024 and always answered 0xbadf5040); SSDT dump file names relative (WRITE FAILED (create failed) before); index newlines; and the POST-GOP verdict no longer compares Turing against 0xffffffff (the shipped Turing chain re-locks on purpose and ends at 0xffffffef).
Adds
- POWER: card PMCSR = … and POWER: root port BB:DD.F VEN… PMCSR = … in ssdiag 0.6.8, with the D-state decoded and a verdict: both D0 = the power path did not engage, a lost unlock is write-type; port D3 = the PR3/_OFF path has run (PERST# asserted, rail off) and the unlock cannot survive it; card D3 with the port in D0 = card-side D3hot only, rail intact. Read-only — the PowerState bits are never written. It is a snapshot, not a history: D0 does not prove the port never went to D3 this boot, and the canary owns that axis.
- CANARY-PRE / CANARY-SET: the loader writes two magic words into the PBUS software-scratch words 0x1480/0x1484 (indices 0x20/0x21) after the FLR verify and reads them back at once. Any FLR, SBR or rail-off clears that block; no shipping driver writes this pair (the driver only uses index 0x1E), so it witnesses exactly the events we are chasing.
- CANARY=<state>(a/b) in POST-GOP and in HANDOFF-OUT, plus CANARY_A/B and a CANARY: verdict line in ssdiag 0.6.7+.
- Verdicts are only meaningful paired with FEAT: canary intact + FEAT gone = a plain write, no reset; canary cleared + FEAT intact = an FLR/SBR-class reset the AON island survived; both gone = rail-off / full-chip reset.
- \tx-no-canary.txt on the USB root skips the whole group (A/B axis).
- PMCBOOT0 is logged as the read-path control.
- \tx-acpi-dump.txt on the USB root dumps the firmware's DSDT and every SSDT to \acpi\ with a keyword index (PR3, ON, OFF, PRST, _ADR, …) for offline iasl -d. Runs before the confirm, needs no GPU, pure read-only. This is how a "who owns the card's power" claim gets settled from the platform AML instead of argued.
- ssdiag 0.6.8: the POWER: section above; discovery still tries the lower log-config priorities and falls back to a config-space BAR0 read when PnP has no resources (safe mode / unstarted device, where 0.6.5 wrongly reported "no NVIDIA adapter"), and prints per-device skip reasons when discovery comes back empty.
- POST-GOP: FEAT_PLM=… SS0=… SS1=… at the handoff tail, after the console rebuild and before StartImage bootmgfw — splits "the card's own GOP/devinit or the firmware wiped it" from "the Windows driver pass did" (#164). Keep the default native-GOP path so the sample points at the right writer.
- HANDOFF-VGA-FINAL bridge routing dump after the VGA restore — the previous HANDOFF-PCI dump ran before the rebuild, so the bridgectl bits Windows actually inherits were never recorded (#165).
Assets: BOOTX64.EFI.2.1.11-diag1 (SHA-256 a85e282be230eb1413fe6718930a343103381b8858893d18f6b98777c738f107), ssdiag.exe.0.6.8 (SHA-256 06f274ccc96bb17b37f8c9c1bb61f4b1d9d864f067b97531ba8f257c75abe3ce, run elevated), and SHA256SUMS.2.1.11-diag1. Build the EFI file into the USB root exactly as for v2.1.10 (make-usbroot.ps1), replacing BOOTX64.EFI.
Protocol: docs/diag-2.1.11-diag1-protocol.md — §3 covers the canary, §5 the ACPI dump, §5b the POWER: reading, §4 the uninitialized-card false negative, §1b the within-OS split for #164, §2 the \tx-select-bdf.txt per-card isolation for #165.
Unlock semantics, payloads and FLR order are unchanged from v2.1.10; results from this build are directly comparable.
v2.1.10 — native-GOP console rebuild becomes the default2026-11-03
v2.1.10 — native-GOP becomes the default console rebuild
---
Attachment replaced in place (2026-11-08). The pinned CachyOS archive dropped nvidia-utils-610.43.03 from its rolling window (#212), which killed every make-usbroot script shipped in this release. The make-usbroot.ps1 / make-usbroot.sh inside the zip (and SHA256SUMS.txt) now carry the fix from main (e3fa6f4): the nvidia-utils input is a source group — Arch Linux Archive → first-party NVIDIA .run → CachyOS cache — with automatic fallback, -UtilsSource / --utils-source to pin one transport, and the same pinned GSP-firmware hashes (all three containers verified bit-identical). Everything else in the zip is byte-unchanged. If you downloaded this release before 2026-11-08, grab the zip again.
附件已原地替换(2026-11-08):CachyOS 归档对每个包只保留约三个版本的滚动窗口,钉定的 nvidia-utils-610.43.03 已被挤出窗口(HTTP 404,#212),本 release 内置的制盘脚本随之失效。zip 内的 make-usbroot.ps1 / make-usbroot.sh 与 SHA256SUMS.txt 已替换为 main 上的修复版(e3fa6f4):nvidia-utils 输入改为源组自动回退——Arch 官方归档(永久保留,版本不清除)→ 第一方 NVIDIA .run 安装器(XFree86 树只增不删)→ CachyOS 旧容器(仅作字节缓存还魂通道)——并可用 -UtilsSource / --utils-source 钉定单一传输。GSP 固件的钉定 SHA-256 不变:三个容器内含固件逐位一致,已交叉验证。zip 其余文件逐字节未动。2026-11-08 之前下载过本 release 的用户请重新下载 zip。
---
Native-GOP console rebuild is now the default on every supported GPU: after the transaction, the loader restarts the card's own GOP driver from its own shadowed option ROM (extract → firmware EFI_DECOMPRESS → explicit Driver-Binding Supported/Start on the target only → GOP-child verification → console rebound on that child). The firmware ConnectController reconnect is no longer in the default path at all.
Credit: the method is not ours. Extracting the card's own stock GOP from its shadowed Option ROM and starting it explicitly under CSM was contributed by @seti9585 in #107 — their own AD103 implementation proved the mechanism and their source review caught the EFI_DECOMPRESS ABI bug in our first version. Their 164608-byte known-answer — identical on the #128 B760/AD103 machine before any fix — was the cross-machine validation anchor for the extraction chain, and #105 (AD104/H670, four lit runs, no 0x193) is the first independent, cross-card confirmation. This release is the direct product of their work.
Why: the firmware reconnect hard-hangs on CSM-enabled boards — five platforms field-recorded (#106 B660, #107 B550, #133 B350, #134 B450, #157 Z390). The native-GOP path's worst case is a graceful fall-through to bootmgfw that still boots Windows; four bench rounds on Z170 + two field machines (B760/AD103, TUF A15/GA107) all reached Windows through every outcome of the new path. Design axiom: a wrong reconnect is worse than none — on any failure (extraction, decompression, binding, child verification) the console stays however boot left it and bootmgfw self-establishes. The old reconnect path remains available as an opt-out: put \tx-reconnect-legacy.txt (empty file) at the stick root.
Observed second effect — unlocked cards, specific outputs, multi-monitor: on top of the CSM story, the native-GOP rebuild was observed to close a second, unrelated failure that this release therefore also retires: after an unlock, a black screen during boot (and a broken Windows session behind it) whenever the card's own output was a specific connector or more than one display. Field evidence, #105: RTX 4070 Ti / AD104 on an H670, three heads (DP1 + DP + HDMI) — v2.1.8 went black on the handoff with Windows alive behind it and a dxgkrnl LiveKernelEvent 0x193 live dump; a single head on the non-primary DP2 also went black, and a monitor hot-plug killed the display for good. On v2.1.10 the identical layouts all boot with every monitor lit, no nvlddmkm events, no 0x193, and the hot-plug just works; the same holds in unlock mode with the four writes and the post-FLR AON readback intact. The reporter's former workaround set (one head on DP1 only, other monitors on a second card, no hot-plug) is no longer needed. Mechanism as far as it is evidenced: the firmware ConnectController left the display engine in its post-FLR state without re-running link training and mode set, so the OS inherited a half-dead pipeline and its recovery watchdog fired; the card's own GOP driver performs the full initialization. Scope caveat: observed so far on two cards on two platforms (AD104/H670 and AD103/B550-F), replay and unlock modes — treat it as a property of this failure family, not as a blanket multi-head guarantee.
Observed third effect — laptop internal panels, brightness (DDC-CI): the same rebuild retired a third failure, reported against the internal panel of some Legion laptops. On one R7000 2024, after an unlock the built-in panel's backlight went dark and DDC-CI brightness control stopped responding; other Legion R7000/R9000 models did not show it. The legacy path reproduced it on the first attempt, and v2.1.10 fixed the same machine — the reporter's words: screen normal, perfect unlock. Mechanism as far as it is evidenced: the legacy path's forced DisconnectController/ConnectController lands in the middle of DP link negotiation on the internal link, and the backlight / DDC-CI channel rides that link, so the OS inherits an untrained link and backlight writes go nowhere. The card's own GOP driver re-runs link training and mode set, so Windows receives a negotiated link — the same "finish the initialization the firmware abandoned" that closes the second effect. Scope caveat: one machine, one panel so far (reported on some Legion SKUs only) — treat it as a property of this failure family, not as a guarantee for every laptop panel. On such machines, opting back into \tx-reconnect-legacy.txt is expected to bring the brightness failure back.
Observed fourth effect — handoff BSOD and virtual display (desktop, TU104): the same rebuild also retires a fourth failure, from the maintainer's own machine (field-confirmed 2026-11-09, issue closed): RTX 2080 SUPER (TU104 1e81) on Z370 (iGPU on, single dGPU head, Windows 10 with GSP) — in the v2.1.9 era the handoff went black and Windows entered an APCINDEXMISMATCH BSOD loop starting ~42 s in (#55; that report predates the native-GOP default and was filed under the "handoff integrity" #35/#61 lineage). On the new build (native-GOP default), the same machine shows neither shape anymore, and the virtual display (software virtual display adapter) form was also retired by the same rebuild — both shapes died with the half-dead pipeline; for the virtual display part only the co-disappearance is evidenced, not a separate mechanism. The report's other dangling reading (a relock value of 0xffffffef) was separately confirmed as a known-good shape (see cold-values / FAQ). Scope caveat: one machine so far — treat it as a property of this failure family.
What else is in
- Boot-option resolution rebuilt (issues #128/#142): dual-ESP machines chained the wrong drive's leftover bootmgfw with a stale BCD (0xc000000f/0xc0000225 — two field machines). The loader now walks the firmware's own BootOrder, skips its own USB entry and its own volume, and chains the loader the firmware itself would boot — resolving full-path FILE options, short-form HD options (partition-identity match, EDK2 BmExpandPartitionDevicePath semantics), bare-FILE short forms, and bulk device scopes (the common real-firmware Windows shape). Operator pins (chain-target.txt) still take precedence; the enumeration-order scan remains the last-resort fallback. Every chain log now