Follow-ups left open by the RTL8852CE PCIe port (#468)

未關閉
#469 1 則留言 0 個 reaction 已指派 0 人 在 GitHub 檢視

維護者通常 1 天內回覆

還沒有人認領這個 Issue。

評估

難度
5/5
預估耗時
一週以上
新手友好度
15/100
Issue 類型
缺陷
描述清晰度
需要釐清
活躍度
活躍
技術堆疊
c
領域
backend

研究方向

This issue groups several independent follow-ups rather than one bounded task. Start with the relevant entry points it names: PcieDmaAx, PcieDma88xx, PcieTransport::Open, tests/pcie_rx_smoke.py, and the demo used by the receiver harness. First determine which item is still actionable; done criteria depend on that investigation, and the lab-box hang remains undiagnosed.

由索引模型根據 Issue 內容生成。

描述

enhancement

Open items from #468 (RTL8852CE on the Kestrel HAL over vfio-PCIe), none of them blocking the merge:

  • MSI on the AX plane. PcieDmaAx polls the RXQ/RPQ hardware indices every 200 µs. The 88xx plane already wires the vfio MSI eventfd; the AX equivalent is R_AX_HAXI_HIMR00 (RXDMA | RPQDMA | RDU) plus the indirect R_AX_PCIE_HIMR00_V1 HCI_AXIDMA bit (rtw89 rtw89_pci_enable_intr_v1 / recognize_intrs_v1). Measure the CPU cost of polling first; it is only worth doing if a receiver on a small host shows it.
  • Card 05:00.0 TX delivery. On the lab box the E-key card loses ~30% of injected frames to either witness (1383/2000 and 1489/2000) while 09:00.0 delivers 1980/2000 with the same code. Confirm with a reseat / antenna swap / SDR duty read before anything in the driver is suspected; if it follows the slot rather than the module, look at the ASPM/L1 state of that root port.
  • The 8821CE plane was moved, not re-run. PcieDma88xx is a byte-identical move of the old ring code, but no 8821CE is on the box. Run tests/pcie_rx_smoke.py on the Radxa 8821CE once.
  • RTL8852BE. Refused at the factory and in PcieTransport::Open: its PCIe power sequence (mac_pwr_on_nic_pcie_8852b) and the single-BDRAM ring table (rtw89_bd_ram_table_single, non-V1 register set, 8-byte addr_info) are unported. The plane would need the 8852B register bank as a second table.
  • CAV-cut pre-init rows. l12_vmain / gen2_force_ib (rtw89, CAV only) are not transcribed; the lab modules are cut 1. Add them if a CAV module ever shows up.
  • timeout -s INT does not stop the PCIe rxdemo while kill -INT <pid> does (clean exit within 5 s). Understand why — a signal-delivery or process-group effect in the demo, not the driver — since every harness that uses timeout on the PCIe receiver silently measures a hard kill.
  • Lab box hang. During a hung-rxdemo investigation the box went unreachable (pings, sshd pre-auth timeouts) and needed a power cycle; the previous boot's kernel log shows nothing at the hang time (only the pre-fix CmpltAbrt AERs and one DMAR fault, both since fixed). Cause open. A serial console or PDU on that node would have turned this into a diagnosis.
  • MMIO alignment doctrine. The transport now splits misaligned accesses; the HALs still issue 32-bit RMWs on 16-bit registers (R_AX_SYS_FUNC_EN at 0x0002 is the canonical one). Consider a debug-build counter of split accesses so new ones are visible.

🤖 Generated with Claude Code

主要語言
C
星號
56
分支
29
平均合併
3 小時 43 分鐘
30 天內合併 PR
75

環境準備

這個專案沒有提供開發容器、Dockerfile 或貢獻指南,環境需要你自己搭建:先看它的 README,通用步驟見我們的新手貢獻指南。

從這裡開始

  1. 先讀完整個 Issue,再讀專案的貢獻指南。
  2. 在 Issue 下留言說明你要接手 —— 這能避免兩個人做同樣的事。
  3. Fork 儲存庫,在一個分支上完成修改。
  4. 送出 Pull Request,並在描述裡引用這個 Issue 編號。

OpenIPC/devourer 的其他 Issue

查看 OpenIPC/devourer 的全部 Issue

相似的 Issue

更多 C Issue

把新 issue 寄到你的電子郵件信箱

精選適合新手參與的 GitHub issue 摘要。