Claude review: gpu: nova-core: run unload sequence upon unbinding

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

Overall Series Review

Subject: gpu: nova-core: run unload sequence upon unbinding
Author: Alexandre Courbot <acourbot@nvidia.com>
Patches: 10
Reviewed: 2026-04-28T15:01:46.459272

---

This is a well-structured 8-patch series that implements GSP unload for nova-core, allowing the driver to properly clean up the GPU when unbinding rather than leaving WPR2 in place and requiring a PCI reset before rebinding. The series is logically decomposed: patches 1-3 are preparatory cleanup, patch 4 adds the GSP shutdown command, patch 5 refactors booter code, patches 6-7 introduce a HAL for chipset-specific boot/unload paths, and patch 8 wires up the full unload sequence (Booter Unloader + FWSEC-SB).

The approach of preparing the unload firmware bundle at probe time (since the filesystem may not be available at unbind) is sound and well-motivated. The HAL introduction is a reasonable forward-looking design for Hopper/Blackwell support.

A few concerns worth raising:

1. **Patch 6 reorders `SetSystemInfo`/`SetRegistry` commands to after RISC-V becomes active** — this is a behavioral change that deserves explicit justification beyond "shuffle code," since previously these were enqueued before the GSP falcon boot and now they're sent after.

2. **Patch 8 makes unload bundle preparation non-fatal** — if it fails, the driver still loads but won't be able to cleanly unload. This is pragmatic but means the driver can silently enter a degraded state.

3. **Pin projection in patch 8** — `self.as_mut().project().unload` assumes `#[pin_data]` generates a `project()` method. This should work with kernel pin-init infrastructure, but is worth verifying.

---

---
Generated by Claude Code Patch Reviewer