From mboxrd@z Thu Jan  1 00:00:00 1970
From: Claude Code Review Bot <claude-review@example.com>
To: dri-devel-reviews@example.com
Subject: Claude review: drm/ttm: Fix GPU MM stats during pool shrinking
Date: Tue, 05 May 2026 08:58:24 +1000
Message-ID: <review-patch1-20260502065338.2720646-1-matthew.brost@intel.com>
In-Reply-To: <20260502065338.2720646-1-matthew.brost@intel.com>
References: <20260502065338.2720646-1-matthew.brost@intel.com> <20260502065338.2720646-1-matthew.brost@intel.com>
X-Mailer: Claude Code Patch Reviewer
Content-Type: text/plain; charset="utf-8"
Content-Transfer-Encoding: quoted-printable
MIME-Version: 1.0

Patch Review

**The helper function (lines +105..+111):**

```c
static void __free_pages_gpu_account(struct page *p, unsigned int order,
				     bool reclaim)
{
	mod_lruvec_page_state(p, reclaim ? NR_GPU_RECLAIM : NR_GPU_ACTIVE,
			      -(1 << order));
	__free_pages(p, order);
}
```

This correctly mirrors the pattern already used in `ttm_pool_free_page()`. =
The ordering (stat update before `__free_pages`) matches the existing code =
and is safe since `mod_lruvec_page_state` only needs the page struct to loo=
k up the lruvec, not the page contents.

**Refactor in `ttm_pool_free_page()` (lines -120..+123):**

Clean replacement of the inline `mod_lruvec_page_state` + `__free_pages` wi=
th the new helper. No functional change. Good.

**Fix in `ttm_pool_restore_commit()` (line +132):**

```c
__free_pages_gpu_account(p, 0, false);
```

Correct. This is the restore path where old retained pages from `tt->pages[=
]` are copied and freed. These pages were allocated via `ttm_pool_alloc_pag=
e()` which marks them `NR_GPU_ACTIVE` (line 175 of current tree). They were=
 never moved to the reclaim pool, so `reclaim=3Dfalse` is right. The order =
is 0 because `ttm_pool_split_for_swap()` was called just above.

**Fix in `ttm_pool_backup()` purge path (line +141):**

```c
__free_pages_gpu_account(page, order, false);
```

Correct. In the purge path, pages from `tt->pages[]` are freed directly at =
their original order (no split). These are `NR_GPU_ACTIVE` pages. `reclaim=
=3Dfalse` is correct. The `page->private =3D 0` reset on the line above cle=
ars the order stored in private, and the helper reads the order from the pa=
rameter rather than from `page->private`, so ordering is fine.

**Fix in `ttm_pool_backup()` swap path (line +150):**

```c
__free_pages_gpu_account(page, 0, false);
```

Replacing `put_page(page)` =E2=80=94 this is the v2 change per Kenneth's fe=
edback. After `ttm_pool_split_for_swap()`, these are individual order-0 non=
-compound pages, so `__free_pages(p, 0)` is functionally equivalent to `put=
_page(page)` here (both do `put_page_testzero` + free). `reclaim=3Dfalse` i=
s correct; these pages are `NR_GPU_ACTIVE`.

**Minor observation:** The `reclaim` parameter is always `false` at all thr=
ee new call sites. The only caller that passes `true` is the existing `ttm_=
pool_free_page()` code path. This is expected =E2=80=94 `ttm_pool_backup()`=
 returns `-EBUSY` for DMA alloc pools, and pages in `tt->pages[]` are never=
 in the reclaim pool (they'd need to go through `ttm_pool_type_add()` for t=
hat). So the `reclaim` parameter is justified by the `ttm_pool_free_page()`=
 usage, but worth noting that the *new* sites are all `NR_GPU_ACTIVE` decre=
ments.

**No issues found.** The patch is a correct and minimal fix for a real acco=
unting bug. The Fixes tag correctly references the commit that introduced t=
he GPU MM stats.

Reviewed-by: Dave Airlie <airlied@gmail.com>

---
Generated by Claude Code Patch Reviewer