**rc6 的 SLOT-REVIVE 在 #276 机器上 declined at parent matching(死槽 bus 2 不在根口当前 [0,0] 范围内——固件把 PEG 口的总线寄存器清零了;PORT-DIAG boot 1 裁决:DLLLA=1 Gen3 x16 链路训练着、D0、无 LinkDisable、secbus=0x00)。rc7 两处修复:
- 父口匹配路径化:固件 device path 前缀匹配优先(编号无关),bus 包含判降级为路径不可读时的 fallback;
- 重编号族从 decline 升级为行动杠杆:把死槽记录的总线号写回父口 sec/sub(读回验证)→ reprobe——若 vendor 读回干净 = 设备只是被路由隔离、从未断电(config 内容含 BARs 完好),无需复位直接保留候选、同 boot 交易;仍 0xffff 才落 SBR 阶梯(此时 config TLP 已恢复路由、杠杆可达)。
给 #276 报告者:直接替换 U 盘 BOOTX64.EFI(TX 版 sha 8c57db7c…200d),markers 保持原样再 boot 一轮。预期:SLOT-REVIVE 按 ROM 身份识别 02:00.0(N 卡)→ bus 恢复 → vendor 读回 0x10de → 同 boot 交易全链 → 系统内 ssdiag 裁决。(只读负控版随行);zip/ssdiag 沿用 rc6(未变)。**
English summary
v2.1.12-rc7 (pre-release) — the #276 renumbering-family fix: path-based parent matching + a bus-restore-first lever.
rc6's SLOT-REVIVE declined at parent matching on the #276 machine (the dead function's bus 2 was not inside the PEG port's current [0,0] range — the firmware ZEROED the PEG port's bus registers; PORT-DIAG boot 1 adjudicated: DLLLA=1, Gen3 x16 trained, D0, no Link Disable, secbus=0x00). Two fixes:
- Path-based parent matching: firmware device-path prefix matching first (numbering-independent); bus-range containment demoted to a fallback for handles with unreadable paths.
- The renumbering family is an active lever now, not a decline: write the dead function's recorded bus back into the port's secondary/subordinate registers (readback-verified), then reprobe. If the vendor reads clean, the device was ROUTING-ISOLATED, never dead — its config space content (BARs included) survived intact, no reset and no bookkeeping restore is needed, the slot stays in the candidate list and the transaction runs same-boot. If it still reads 0xffff, the ladder falls through to the SBR levers, now reachable because config TLPs route again.
For the #276 reporter: replace the stick's BOOTX64.EFI with the TX build (sha 8c57db7c…200d), keep the markers as they are, and boot once. Expected log: SLOT-REVIVE identifies 02:00.0 via the ROM gate (NVIDIA) → bus registers restored → vendor reads 0x10de → the full transaction targets the card at its original bus-2 address → in-OS ssdiag completes the record. Running the transaction at the card's original bus-2 location is itself the proof of the renumbering diagnosis. BOOTX64.EFI.readonly.2.1.12-rc7 (read-only control, sha 15130fc2…176f) ships alongside; the zip and ssdiag are unchanged from rc6.
Mechanism note: this family is bus-number hiding — the firmware strips the bridge's routing label while deliberately keeping the display face alive (link trained, VGA forwarding set, memory windows routing: the console kept painting; scanout never stopped, which also proves the card was never reset). It is spec-legal (the PCI register default is 0/0, and UEFI imposes no handoff invariant on bridge numbering) but assumes the next enumerator will rediscover the device on its own — Windows does, our loader now does too.