Claude review: drm/v3d: Introduce Runtime Power Management

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

Patch Review

**Status: Looks good overall with minor observations.**

This is the main patch. The implementation follows standard runtime PM patterns well.

**v3d_power.c:** The suspend/resume callbacks are clean — disable IRQs, power off SMS, disable clock on suspend; reverse on resume. The `v3d_resume_sms`/`v3d_suspend_sms` are direct moves from the old `v3d_idle_sms`/`v3d_power_off_sms` with appropriate renames.

**Probe changes:** The transition from direct `clk_prepare_enable()` to `pm_runtime_resume_and_get()` is correct. The `CONFIG_PM` guard for calling `v3d_power_resume()` manually is the right pattern. The 100ms autosuspend delay (bumped from 50ms in v7 due to RPi 4 issues) seems reasonable.

**Job submission (v3d_submit.c):** Acquiring a PM reference in `v3d_job_init()` for non-CPU jobs and releasing it in `v3d_job_free()` is a clean approach — it keeps the GPU powered while jobs are in flight. The `has_pm_ref` flag prevents double-put.

**MMU flush (v3d_mmu.c):** Using `pm_runtime_get_if_active()` to skip the flush when the device is already suspended is correct — if the GPU is off, the page tables will be reloaded on resume anyway.

**Perfmon (v3d_perfmon.c):** Similarly using `pm_runtime_get_if_active()` in `v3d_perfmon_start()` and `v3d_perfmon_stop()` is correct. In `v3d_perfmon_stop()`, the `out_clear` label that clears `active_perfmon` even when the device is suspended makes sense — we still need to track that the perfmon is no longer active.

**Debugfs:** Adding `v3d_pm_runtime_get()`/`v3d_pm_runtime_put()` around HW register reads in debugfs is correct.

**Minor observation on v3d_platform_drm_remove():**

```c
	pm_runtime_suspend(dev);

	/* If PM is disabled, we need to call v3d_power_suspend() manually. */
	if (!IS_ENABLED(CONFIG_PM))
		v3d_power_suspend(dev);
```

Using `pm_runtime_suspend()` directly is slightly unusual — most drivers use `pm_runtime_disable()` + `pm_runtime_set_suspended()` or rely on `devm_pm_runtime_enable()` cleanup. However since `devm_pm_runtime_enable()` is used, the disable will happen via devm cleanup. The explicit `pm_runtime_suspend()` forces the device to suspend state before tearing down resources, which is fine.

**One question:** In `v3d_get_param_ioctl()`, the PM get/put wraps the register reads, but for the `DRM_V3D_PARAM_SUPPORTS_*` cases (which return constant values), the function returns early before hitting the PM calls. This is correct — no HW access needed for those params.

Overall this is a well-tested series (v8!) with proper review tags. The code is clean and follows kernel runtime PM conventions correctly.

---
Generated by Claude Code Patch Reviewer