Claude review: drm/ttm: Split cgroup charge and resource allocation

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

Patch Review

This is the most complex refactoring patch. It splits `ttm_resource_alloc()` into `ttm_resource_try_charge()` + `ttm_resource_alloc()`, keeping the charge alive across eviction retries.

**Key design**: `charge_pool` is charged once, and ownership transfers to the `ttm_resource` on success (set to NULL in alloc_state). On failure to place, the charge is explicitly undone with `dmem_cgroup_uncharge()`.

**The `in_evict` flag**: prevents recursive eviction — `may_evict` is forced false during the evict callback. This is important since `ttm_bo_alloc_at_place` is now called from `ttm_bo_evict_cb`.

**Review of uncharge paths in `ttm_bo_alloc_resource()`**:
- `-ENOSPC` path: uncharges + puts limit_pool + `continue` ✓
- `-EBUSY` path with evict failure: uncharges + `continue` or `return` ✓  
- Other error path: uncharges + puts limit_pool + `return` ✓
- Success path: charge transferred to resource ✓

This looks correct. The `ttm_resource_alloc()` signature change from `**ret_limit_pool` to `*charge_pool` is clean.

**One concern**: In the evict callback path (`ttm_bo_evict_cb`), `ttm_bo_alloc_at_place` is called with `force_space=false`. Since `charge_pool` is already set (charged earlier), the `if (!alloc_state->charge_pool)` block is skipped, going straight to `ttm_resource_alloc()`. This is correct — we don't want to re-charge. But `may_evict` will be false (both `in_evict` prevents it and `force_space` is false), so allocation failure returns the raw error. Since `lret` in the callback treats `0` as success and negative as "keep trying", this should be fine.

---
Generated by Claude Code Patch Reviewer