jetson-customize-mgbe

par nvidia

Activer le Jetson Thor 25G/10G/1G MGBE QSFP via overlay kernel-DT. Ne PAS utiliser pour l'allocation de voies UPHY ou les modifications ODMDATA.

npx skills add https://github.com/nvidia/skills --skill jetson-customize-mgbe

Customize MGBE / 25G QSFP

Overview

Thor T264 exposes mgbe0..mgbe3. On a custom carrier, the 25G QSFP cage (or 10G / 1G fiber path) is wired to one of them through SerDes — with or without an external MDIO PHY in front of the cage. This skill renders the kernel-DT overlay that pairs the BPMP allocation with kernel-side status="okay" + PHY plumbing on &mgbeN.

Out of scope:

  • UPHY lane allocation — owned by /jetson-customize-uphy. Refuse if the chosen uphy1-config-N doesn't allocate the target MGBE.
  • All ODMDATA tokens (mgbeN-speed-*, sub-node mgbeN_status=*) — owned by /jetson-customize-uphy in its single atomic ODMDATA commit. This skill MUST NOT touch ODMDATA=.

Output is one commit to the composite custom overlay .dts in the bsp_sources/ hardware repo. /jetson-build-source compiles the composite to .dtbo and owns its Makefile + flash-conf registration.

When to invoke

  • The user says "enable 25G", "configure QSFP", "set MGBE PHY mode", "wire MGBE to QSFP", or asks to bring up a 10G / 1G fiber path.
  • Cold boot succeeds but ip link show mgbe<N> reports state DOWN or NO-CARRIER on the configured controller, OR the controller never appears at all.
  • jetson-customize-uphy ran with uphy1-config-8 (or another config allocating MGBE) and you now need to bring up the per-controller side.

Prerequisites:

  • Active profile selected with reference_devkit: (Thor) + custom_carrier: blocks.
  • <source.root_path>/Linux_for_Tegra/.git exists (/jetson-init-source).
  • /jetson-derive-carrier has run — carrier flash-conf fork is in the overlay tracker.
  • /jetson-customize-uphy chose a UPHY config that allocates the target MGBE controller's lanes (uphy1-config-8 on Thor for MGBE0..3 25G).
  • Source-of-truth docs registered or supplied at prompt: Adaptation Guide, Module Design Guide, SoC TRM.
  • When custom_carrier: is present, both documents.custom_carrier_schematic AND documents.custom_carrier_pinmux_xls are REQUIRED. Refuse the run if either is missing — MGBE routing on a custom carrier cannot be guessed. Reference-devkit-only profiles skip this check.
  • dtc on PATH.

Procedure

See references/procedure.md for the full step-by-step procedure (Steps 1–8). Summary:

  1. Resolve active target + documents. Validate active profile, custom_carrier, overlay tracker; locate the relevant Adaptation Guide chapter and pinmap.
  2. Per-controller question loop. AskUserQuestion driven by questions.json (controller, phy_mode, attach kind, I²C bus/addr, reset GPIO, compatible_list).
  3. Derive max-speed from phy_mode. Decompile the BPMP DTB to pick sub-node vs top-level token grammar; cite the inspection in notes.
  4. Verify HSIO pins + auto-fix. Run pin_verifier.py for MDC/MDIO/RESET/INT; surface mismatches and route to /jetson-customize-pinmux.
  5. (no ODMDATA edits.) MGBE ODMDATA tokens are emitted by /jetson-customize-uphy. Step 5 only records the BPMP DTB token-form inspection (sub-node vs top-level) in notes[] for audit.
  6. Append composite-overlay fragments. Write one fragment per controller into the composite custom overlay .dts; obey the /* custom-bsp: mgbe:mgbe... */ marker contract; run the cpp/dtc/fdtoverlay pre-flight.
  7. (Reserved.) Sibling-skill ordering / cross-cutting validation.
  8. Run-state sidecar + summary + next-step chain. Write <profile-stem>.jetson-customize-mgbe.json and emit the one-line + table summary, then drive the downstream chain via sequential AskUserQuestion prompts per references/procedure.md Step 8. Never substitute a printed "Next step: …" line for the prompts.

Gotchas

  • Stock Thor BPMP DTB has no /mgbe/mgbe@N subtree — only mgbe<N>-speed under /uphy. The mgbe<N>_status=disabled sub-node token is silently rejected on these releases; the whole ODMDATA line is then dropped at flash time. Always decompile BPMP DTB (Step 3) before emitting; use the top-level dashed form (mgbe<N>-speed-del to remove, mgbe<N>-speed-25G to set) when the sub-node isn't there. Same wrong-form failure surface as jetson-customize-uphy.
  • mdio child needs both #address-cells = <1> AND #size-cells = <0> when phy_attach_kind=="phy". Missing either → kernel rejects phy@<addr> reg property at probe; MGBE never comes up.
  • Overlay root compatible must intersect live DT compatible. UEFI plugin-manager filters by compatible match. A mismatched overlay is silently skipped — flash succeeds, MGBE stays disabled, no error in dmesg. Always sanity-check against /proc/device-tree/compatible on a booted reference DUT.
  • OVERLAY_DTB_FILE ordering is jetson-build-source's problem, not this skill's. This skill never touches the carrier flash conf. The composite custom overlay is registered (by /jetson-build-source Step 5.0a) AFTER the platform *-dynamic.dtbo, which is the correct ordering. If you find yourself appending OVERLAY_DTB_FILE+= in this skill, you're duplicating ownership — stop, and let the build skill do it.
  • UPHY lane allocation is jetson-customize-uphy's job. If the chosen uphy1-config-N doesn't allocate lanes for the target MGBE controller, BL31 SError (fmon_update_config: detected fault 0x80) on cold boot. Always run /jetson-customize-uphy first; cite the chosen uphy1-config-N in this skill's run-summary notes[].
  • Don't disable a stock-okay controller via ODMDATA alone. Same rule as jetson-customize-uphy: mgbe<N>_status=disabled for a controller that's already disabled in BPMP DTB is a no-op the parser may treat as ambiguous → drops the rest of the ODMDATA line. Disable via the kernel-DT overlay (status="disabled") only.
  • Don't touch the upstream BSP at <bsp_image.root_path>. All edits land in the overlay tracker / bsp_sources mono-repo under the pristine + customization commit pattern.
  • JSON sidecar is structured state, not authoritative. Same caveat as jetson-customize-uphy: ODMDATA + overlay .dts + two git commits are the device-facing outputs; the sidecar is for tooling and idempotency only.

References

Plus de skills de nvidia

compileiq-debug
nvidia
Utilisez quand quelque chose ne va pas : Search() bloque, toutes les évaluations retournent INVALID_SCORE, les scores ne s'améliorent pas, chaque configuration retourne le même nombre, erreurs ptxas…
create-github-pr
nvidia
Créer des pull requests GitHub en utilisant l'interface en ligne de commande gh. Utiliser lorsque l'utilisateur souhaite créer une nouvelle PR, soumettre du code pour révision, ou ouvrir une pull request. Mots-clés de déclenchement -…
nemoclaw-maintainer-cross-issue-sweep
nvidia
Analyse les autres problèmes ouverts pour trouver ceux qu’une PR donnée pourrait également corriger ou casser accidentellement. Génère des opportunités de correctifs adjacents et des risques de contradiction avec fichier:ligne…
fhir-basics
nvidia
Apprend aux agents comment fonctionnent les API FHIR R4, quelles ressources sont disponibles, comment les interroger avec des paramètres de recherche, et comment analyser correctement tous les formats de réponse…
compileiq-validate-result
nvidia
Utiliser APRÈS qu'une recherche soit terminée et AVANT de réclamer un accélérateur ou d'expédier un ACF. Charge le CSV dump_results, extrait les K meilleurs candidats (mono-objectif)…
changelog-audit
nvidia
Auditer le CHANGELOG.md de Warp avant une publication : récupérer les entrées perdues, trier par impact utilisateur, affiner le langage des entrées, ajuster les retours à la ligne et (en mode branche de publication) mettre à jour la comparaison…
maintain-dynamic-plugins
nvidia
Maintenir les chargeurs de plugins dynamiques NeMo Relay, les manifestes, les SDK natifs Rust, le protocole worker gRPC, le SDK worker Python, la documentation, les tests et la couverture du workflow de publication
dgx-diagnose
nvidia
Diagnostiquer les problèmes courants du DGX Station GB300 — plantages CUDA, ciblage incorrect du GPU, bugs de conteneur vLLM/SGLang, problèmes d'état MIG, erreurs NVLink/Fabric Manager,…