Claude review: drm/panthor: Reduce dma_fence signalling latency

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

Overall Series Review

Subject: drm/panthor: Reduce dma_fence signalling latency
Author: Boris Brezillon <boris.brezillon@collabora.com>
Patches: 41
Reviewed: 2026-05-16T13:32:44.342336

---

This is a well-structured 11-patch series by Boris Brezillon that moves panthor's JOB and GPU interrupt processing from threaded IRQ handlers / work items into the hard IRQ handler, reducing dma_fence signalling latency by ~5% on rk3588. The series follows a clean progression:

1. **Patches 1-4**: Mechanical refactoring — make IRQ state non-atomic, reorder code, replace macro machinery with inline helpers, and add the hard-handler plumbing.
2. **Patch 5**: Make FW req helpers IRQ-safe (`spinlock_irqsave`).
3. **Patches 6-7**: Prepare the scheduler for IRQ-context event processing — introduce `events_lock`, inline fence signalling, automate CSG IRQ processing at unbind.
4. **Patch 8**: Optimize interrupt usage with on-demand `ack_irq_mask` enablement.
5. **Patches 9-11**: Actually switch JOB and GPU IRQ processing to the hard handler.

The overall approach is sound. The MMU handler is correctly left as a threaded handler since it needs sleepable locks. The lock conversion to `irqsave` variants is methodical and mostly correct. Patches 1-4 already have Reviewed-by tags from Steven Price and Liviu Dudau.

There are a few concerns worth raising, primarily around patch 9's custom IRQ handling logic and the interaction of `events_lock` scope with `group_unbind_locked`.

---

---
Generated by Claude Code Patch Reviewer