Claude review: drm/ttm: Hook up a cgroup-aware reclaim callback for the dmem controller

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

Patch Review

**Verdict: Good with one concern worth discussing.**

The `ttm_bo_evict_cgroup()` function mirrors the multi-pass strategy of `ttm_bo_evict_alloc()`: trylock pass, low-watermark pass, then sleeping-lock passes. The structure is clean.

The new `sleeping_lock` field in `ttm_lru_walk_arg`:

```c
bool sleeping_lock;
```

The change in `__ttm_bo_lru_cursor_next()`:

```c
-		else if (!arg->ticket || arg->ctx->no_wait_gpu || arg->trylock_only)
+		else if ((!arg->ticket && !arg->sleeping_lock) || arg->ctx->no_wait_gpu ||
+			 arg->trylock_only)
```

When `sleeping_lock = true` and `ticket = NULL`, this falls through to `ttm_lru_walk_ticketlock()` which calls `dma_resv_lock(bo->base.resv, NULL)`. Without a WW ticket, there's no deadlock detection. As the commit message notes, this relies on the caller only holding one lock at a time. This is true here since the cgroup reclaim path has no pre-existing `dma_resv` locks. Using `interruptible = true` mitigates hard deadlocks since the process can be signalled, but a theoretical lock-order inversion could still cause a soft deadlock. Acceptable for this use case (admin writing to cgroup files), but should be flagged as a limitation.

The modification to `ttm_bo_evict_cb()` for cgroup drain mode:

```c
+	s64 bo_size = bo->base.size;
...
+	} else {
+		/* Cgroup drain: return bytes freed for byte-denominated progress. */
+		return bo_size;
+	}
```

This captures `bo->base.size` before eviction. `bo->base.size` is the GEM object size which doesn't change during eviction, so the early capture is defensive but harmless. Returning `bo_size` instead of `1` is correct since `ttm_lru_walk_for_evict()` accumulates these values and the caller (`ttm_bo_evict_cgroup`) compares against `target_bytes`.

The comment about `evict_walk->place` being NULL in cgroup drain mode is important:

```c
+	/*
+	 * evict_walk->place is NULL in cgroup drain mode.  Drivers'
+	 * eviction_valuable() callbacks must handle a NULL place, treating it
+	 * as "any placement": the TTM base implementation already does so via
+	 * ttm_resource_intersects().
+	 */
```

I verified that `ttm_resource_intersects()` returns `true` when `place == NULL` (at `ttm_resource.c:450`), so `ttm_bo_eviction_valuable()` correctly handles this. The xe driver's `xe_bo_eviction_valuable()` delegates to `ttm_bo_eviction_valuable()` as its first check and doesn't use `place` directly, so it's also safe without modification. Only amdgpu needs the explicit NULL guard (patch 5).

In `ttm_resource_manager_set_dmem_region()`:

```c
if (!IS_ERR_OR_NULL(region)) {
    man->cg = region;
    dmem_cgroup_region_set_reclaim(region, ...);
}
```

This correctly handles the error/NULL case (v3 fix from Sashiko-bot), leaving `man->cg` at its default (NULL/zero) if registration failed.

---
Generated by Claude Code Patch Reviewer