Claude review: drm/nouveau: nv04 FIFO cleanup + recovery for Tesla

[Claude Code Review Bot <claude-review@example.com>]

Overall Series Review

Subject: drm/nouveau: nv04 FIFO cleanup + recovery for Tesla
Author: Marek Czernohous <mczernohous@gmail.com>
Patches: 3
Reviewed: 2026-05-16T11:34:49.332858

---

This is a well-structured two-patch series adding observability and recovery to the legacy nv04 FIFO path on Tesla GPUs. The cover letter is thorough, the design (Tier-1 channel-kill + Tier-2 sliding-window wedge) is sound, and the code follows existing nouveau patterns (e.g., the `nvkm_chan_get_chid` / `nvkm_chan_error` / `nvkm_chan_put` sequence mirrors Fermi+ paths in gf100.c, ga100.c, gv100.c).

The series is relatively clean. The main issues are:

1. **German comments in the public header** — the `struct nvkm_fifo_wedge` field comments are in German. Kernel code should use English.
2. **No unwedge/reset mechanism** — once `atomic_xchg(&w->wedged, 1)` fires, Tier-2 is permanently disabled for the lifetime of the driver. This is a design point worth an explicit note or a deliberate decision to document.
3. **Wedge init/fini runs for all GPU generations** — `nv04_fifo_wedge_init`/`fini` is called unconditionally from `nvkm_fifo_new_`/`nvkm_fifo_dtor`, even though `nv04_fifo_recover` is only called from the nv04 FIFO interrupt handler. Harmless but slightly untidy.

No blocking issues found. The locking is correct, `schedule_work` from IRQ context is fine, and the workqueue indirection for `drm_dev_wedged_event` (which may sleep via `kobject_uevent_env`) is the right approach.

---

---
Generated by Claude Code Patch Reviewer