nvidia-skill-finder
Use for NVIDIA-related requests where an NVIDIA skill might help, even if the user did not ask for a skill. Trigger on NVIDIA products, hardware, software,…
Per-controller PCIe enable / disable / lanes / link-speed for a Jetson Thor or Orin custom carrier via ODMDATA + kernel-DT overlay. Do NOT use for UPHY lane allocation or endpoint-mode bring-up.
$ npx -y skills add NVIDIA/skills --skill jetson-customize-pcie --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/jetson-customize-pcieContext preview
The summary Claude sees to decide when to auto-load this skill.
Per-controller PCIe enable / disable / lanes / link-speed for a Jetson Thor or Orin custom carrier via ODMDATA + kernel-DT overlay. Do NOT use for UPHY lane allocation or endpoint-mode bring-up.
name: jetson-customize-pcie
description: >-
Per-controller PCIe enable / disable / lanes / link-speed for a
Jetson Thor or Orin custom carrier via ODMDATA + kernel-DT overlay.
Do NOT use for UPHY lane allocation or endpoint-mode bring-up.
version: 0.0.1
license: "Apache-2.0"
metadata:
data-classification: public
author: "Jetson Team"
tags:
- bsp
- phase-2
- io
- pcie
domain: metaPCIe on Tegra264 (Thor, `pcie@C0..C5`) and Tegra234 (Orin, `pcie@C0..C10`) is split across multiple controllers that share the UPHY lane pool with USB3 / MGBE / UFS. Each controller's runtime behavior is determined by **two surfaces, both required**:
| Surface | Target | Authoritative for | |---|---|---| | ODMDATA `pcie@N_status=…` (+ `pcie@N_max-link-speed`, `pcie@N_pcie-mode`, `pcie@N_clk-scheme`, `pcie-cN-endpoint-enable`) | `/pcie/pcie@N` in BPMP DTB | UPHY lane power, refclk gating, controller-side power rails | | Kernel-DT overlay on `&pcieN` | `/bus@0/pcie@<addr>` in kernel DTB | Kernel probe, lane width, link speed, RC/EP mode |
Skipping the kernel overlay on a disable lets the kernel probe a powered-down PHY (link timeouts in dmesg). Skipping the ODMDATA token on a disable leaves BPMP holding the PHY hot.
**Agentic, not table-driven** — no controller table, no `questions.json`. Every controller, lane width, schematic-routed receptacle, and authoritative DT node address is discovered at runtime from the docs + DTB + carrier pinmap.
The output is a **kernel-DT overlay commit only**. Per-controller `fragment@N` blocks are appended to the composite custom overlay `.dts` per [`../../references/bsp-customization-kernel-dtb.md`](../../references/bsp-customization-kernel-dtb.md) and committed to the `bsp_sources/` hardware repo. `/jetson-build-source` compiles the composite to `.dtbo` and owns its Makefile + flash-conf registration.
**This skill MUST NOT edit `ODMDATA="..."`.** All ODMDATA tokens (`pcie@N_status=…`, `pcie@N_max-link-speed`, `pcie@N_pcie-mode`, `pcie@N_clk-scheme`, `pcie-cN-endpoint-enable`, plus the `uphyX-config-N` surface tokens and `UPHY_CONFIG=""` clear) are emitted by `/jetson-customize-uphy` in a single atomic commit on the carrier flash-conf fork. The allocation table this skill consumes from the UPHY sidecar already tells the operator which controllers are `okay` / `disabled` / per-lane sized; this skill only translates that table into kernel-DT overlay fragments and verifies that the overlay agrees with the ODMDATA already committed by `customize-uphy` (consistency check in Step 8 — disagreement is reported, not silently fixed).
num-lanes", "change PCIe link speed", or asks to flip a `pcie@N_status` token.
flash, OR the link trains at the wrong width / speed.
(e.g. switched from `uphy0-config-7` to `uphy0-config-6` enabling PCIe C3); the per-controller side now needs to be brought up.
doesn't probe its PCIe-side companion (rare; XFI configurations).
**Prerequisites:**
(`/jetson-init-source`).
the overlay tracker.
`<workspace>/target-platform/<profile-stem>.jetson-customize-uphy.json` drives the per-controller `enable` decision.
Guide, Module Design Guide, SoC TRM.
`documents.custom_carrier_schematic` AND `documents.custom_carrier_pinmux_xls` are REQUIRED.** Refuse the run if either is missing — routing on a custom carrier cannot be guessed. Reference-devkit-only profiles skip this check.
Full step-by-step walkthrough lives in [`references/procedure.md`](references/procedure.md). High-level flow:
1. Resolve active target + open source-of-truth documents (incl. `<carrier-pinmap>`, `<ref-dtb>`, `<uphy-state>`). Refuse if `<uphy-state>` is missing. 2. Diff PCIe topology — devkit vs custom carrier — by decompiling `<ref-dtb>` and grepping the schematic for `PEX<N>_*` net labels. 3. `AskUserQuestion` (multiSelect) — which controllers to customize. 4. Per-controller verification: pinmap + schematic + `pin_verifier.py` for `PE<N>_CLKREQ_L`, `PE<N>_RST_L`, optional `PE<N>_WAKE_L`. 5. Auto-derive per-controller plan (`enable` from `<uphy-state>`, `lanes` / `speed` from Adaptation Guide, `mode` hard-pinned to `"rc"`) → mandatory confirm-or-customize gate. 6. Append per-controller `fragment@N` blocks (marker `/* custom-bsp: pcie:pcie@<addr> */`) to the composite custom overlay `.dts` in `bsp_sources/`. Pre-flight `dtc` + `fdtoverlay`. Commit via the workflow's preview gate. **Do not edit `ODMDATA`** — `/jetson-customize-uphy` already emitted `pcie@N_status=…`, `pcie@N_max-link-speed`, `pcie@N_pcie-mode`, `pcie@N_clk-scheme`, and `pcie-cN-endpoint-enable` in its single atomic ODMDATA commit. This skill only translates the per-controller plan into kernel-DT overlay fragments. 7. (Step folded into Step 6 — overlay-only emission.) 8. Cross-check ODMDATA vs overlay consistency. On a contradictory row, **stop and ask the user how to recover the two commits. Never run `git reset --hard` autonomously.** 9. Write run-state JSON sidecar at `<workspace>/target-platform/<profile-stem>.jetson-customize-pcie.json`
`AskUserQuestion` prompts per `reference
Official, NVIDIA-verified Agent Skills for Claude Code, Codex, and other coding agents.
Use for NVIDIA-related requests where an NVIDIA skill might help, even if the user did not ask for a skill. Trigger on NVIDIA products, hardware, software,…
Official NVIDIA-authored guidance for NVIDIA cuDF GPU DataFrames, pandas acceleration, dask-cuDF, ETL, joins, groupby, CSV/Parquet I/O, nullable semantics, and…
Use when asked to install, deploy, run, validate, troubleshoot, or stop NVIDIA AI-Q Blueprint infrastructure.
Use when asked to run deep research or AI-Q research through a reachable NVIDIA AI-Q Blueprint backend.
Calibrate a new dataset from live RTSP camera streams via the AutoMagicCalib REST API. Use when the user provides RTSP URLs or asks to calibrate live cameras;…
Run end-to-end calibration on the shipped sample dataset (sdg_08_2_sample_data_010926.zip) against a running AMC microservice. Use when user says 'test sample…