Claude review: drm/panthor: Track the number of mmap on a BO

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

Patch Review

Tracks mmap count using `refcount_t` to determine reclaimability ordering. The pattern used is:

```c
if (!refcount_inc_not_zero(&bo->cmap.mmap_count)) {
    dma_resv_lock(obj->resv, NULL);
    if (!refcount_inc_not_zero(&bo->cmap.mmap_count))
        refcount_set(&bo->cmap.mmap_count, 1);
    dma_resv_unlock(obj->resv);
}
```

This double-check pattern uses the resv lock as the serialization mechanism for the 0→1 transition. The `refcount_set` on the 0→1 path is intentional because `refcount_inc` from zero triggers a WARN with `CONFIG_REFCOUNT_FULL`. The outer `refcount_inc_not_zero` is a fast path to avoid the lock for the common N→N+1 case.

The `panthor_gem_vm_close` uses a similar pattern for the 1→0 transition:

```c
if (refcount_dec_not_one(&bo->cmap.mmap_count))
    goto out;

dma_resv_lock(bo->base.resv, NULL);
if (refcount_dec_and_test(&bo->cmap.mmap_count)) {
    /* Nothing to do, pages are reclaimed lazily. */
}
dma_resv_unlock(bo->base.resv);
```

**Potential issue**: `refcount_dec_not_one()` returns true when the count was decremented (i.e., went from N>1 to N-1). If it returns false (count was 1), the code takes the lock and calls `refcount_dec_and_test()`. But between `refcount_dec_not_one()` returning false and the lock being acquired, another thread could have incremented the refcount (via mmap). In that case `refcount_dec_and_test()` would decrement but not reach zero, which is fine — but the lock was acquired unnecessarily. The correctness holds because the resv lock serializes the critical state transitions.

One nit: `panthor_gem_vm_open` calls `drm_gem_is_imported()` to skip the mmap_count increment for imported BOs. But `panthor_gem_mmap` redirects to `dma_buf_mmap` for imported BOs and returns early, so the mmap_count would never have been incremented in the first place. The check is defensive but slightly misleading.

No blocking issues.

---
Generated by Claude Code Patch Reviewer