Claude review: drm/ttm: Introduce ttm_bo_shrink_kswap_maybe_fragmented()

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

Patch Review

**Overall:** Clean helper that combines the kswapd check, order check, NUMA node check, and zone fragmentation heuristic.

```c
bool ttm_bo_shrink_kswap_maybe_fragmented(int nid, s8 order)
{
    ...
    if (!order)
        return false;
    if (!current_is_kswapd())
        return false;
    if (!numa_valid_node(nid))
        return false;
    ...
    for (; zone_type <= ZONE_NORMAL; ++zone_type) {
        struct zone *zone = &NODE_DATA(nid)->node_zones[zone_type];
        if (zone_maybe_fragmented_in_shrinker(zone))
            return true;
    }
    return false;
}
```

**Concerns:**

- **Logic inversion on `order`:** The function returns `false` when `order == 0`. This means "order-0 reclaim is never considered fragmentation-driven," which makes sense — order-0 allocations can't fragment. But it also means the function will say "maybe fragmented" for *any* non-zero order (e.g., order-1), even if only order-9 allocations are causing the fragmentation. The function doesn't compare `order` against `beneficial_order` — it simply returns true if kswapd is running, order > 0, and the zone looks memory-rich. This seems intentionally broad but could cause the Xe shrinker to skip reclaim even when order-1 allocations genuinely need help.

- **Zone iteration:** The `#if IS_ENABLED(CONFIG_ZONE_DMA32)` preprocessor check starts the loop at `ZONE_DMA32` when configured, otherwise `ZONE_NORMAL`. The loop always goes up to `ZONE_NORMAL` inclusive. On a system with both DMA32 and NORMAL zones, if *either* zone appears fragmented, the function returns true. This means a fragmented DMA32 zone (which GPU drivers often care about for 32-bit DMA addresses) can cause the shrinker to skip reclaim even if NORMAL zone is fine. This seems correct for GPU workloads but worth noting.

- **ZONE_DMA not checked:** Systems with `CONFIG_ZONE_DMA` are not handled — the loop starts at `ZONE_DMA32` or `ZONE_NORMAL`. This is probably fine since GPU allocations typically don't target the legacy DMA zone (ISA DMA, first 16MB), but a brief comment would help.

- **Missing `#include`:** The function uses `zone_maybe_fragmented_in_shrinker()` from `linux/vmstat.h`. I'd verify `ttm_bo_util.c` already includes `linux/vmstat.h` or pulls it in transitively, otherwise this will fail to compile.

- **EXPORT_SYMBOL vs EXPORT_SYMBOL_GPL:** This exports with `EXPORT_SYMBOL` matching `ttm_bo_shrink_avoid_wait` above it. Fine for consistency.

---
Generated by Claude Code Patch Reviewer