Claude review: drm/v3d: Allocate all resources before enabling the clock

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

Patch Review

**Status: Looks good, one minor nit.**

This patch reorders probe so that all resource allocation (reg mapping, reset control, clk get, IRQ init, perfmon init, DMA alloc, gem init) happens before the clock is enabled and HW is accessed. This is the right preparation for runtime PM.

Key changes:
- `v3d_init_hw_state()` and `v3d_mmu_set_page_table()` are moved out of `v3d_gem_init()` and called explicitly after `clk_prepare_enable()`. `v3d_init_hw_state()` is made non-static (exported via header).
- `v3d_irq_init()` is split: registration happens early, but `v3d_irq_enable()` (which touches HW) is deferred until after the clock is on. The interrupt clearing is moved from `v3d_irq_init()` into `v3d_irq_enable()`, which makes sense since it needs HW access.
- The remove order is corrected to be the reverse of init (clock disable before gem_destroy, as the gem layer shouldn't need HW at teardown).

**Minor nit:** In `v3d_gem.c` after moving `v3d_init_hw_state()`, there's a double blank line:

```c
void
v3d_init_hw_state(struct v3d_dev *v3d)
{
	v3d_init_core(v3d, 0);
}


static void
v3d_huge_mnt_init(struct v3d_dev *v3d)
```

One blank line would be sufficient per kernel style.

**One concern about the error paths:** After the reorder, there's a potential issue in the probe error cleanup. The `irq_disable` label calls `v3d_irq_disable(v3d)`, but now falls through to `clk_disable` which calls `v3d_power_off_sms(v3d)` and `clk_disable_unprepare(v3d->clk)`, then to `gem_destroy` and `dma_free`. This ordering looks correct — if `drm_dev_register()` fails, we need to disable IRQs, turn off the clock, then clean up gem/dma. The label ordering in the diff (`irq_disable` → `clk_disable` → `gem_destroy` → `dma_free`) correctly reverses the init sequence.

---
Generated by Claude Code Patch Reviewer