[PATCH] drm/ttm: Fix GPU MM stats during pool shrinking

[PATCH] drm/ttm: Fix GPU MM stats during pool shrinking

[Matthew Brost]

TTM pool shrinking frees pages by calling __free_pages() directly,
which bypasses updates to NR_GPU_ACTIVE and leaves GPU MM accounting
out of sync.

Introduce a helper, __free_pages_gpu_account(), and use it for all page
frees in ttm_pool.c so GPU MM statistics are updated consistently.

Reported-by: Kenneth Crudup <kenny@panix.com>
Fixes: ae80122f3896 ("drm/ttm: use gpu mm stats to track gpu memory allocations. (v4)")
Cc: Christian Koenig <christian.koenig@amd.com>
Cc: Huang Rui <ray.huang@amd.com>
Cc: Matthew Auld <matthew.auld@intel.com>
Cc: David Airlie <airlied@gmail.com>
Cc: dri-devel@lists.freedesktop.org
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
---
 drivers/gpu/drm/ttm/ttm_pool.c | 16 +++++++++++-----
 1 file changed, 11 insertions(+), 5 deletions(-)

diff --git a/drivers/gpu/drm/ttm/ttm_pool.c b/drivers/gpu/drm/ttm/ttm_pool.c
index 26a3689e5fd9..95bbd9328072 100644
--- a/drivers/gpu/drm/ttm/ttm_pool.c
+++ b/drivers/gpu/drm/ttm/ttm_pool.c
@@ -206,6 +206,14 @@ static struct page *ttm_pool_alloc_page(struct ttm_pool *pool, gfp_t gfp_flags,
 	return NULL;
 }
 
+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);
+}
+
 /* Reset the caching and pages of size 1 << order */
 static void ttm_pool_free_page(struct ttm_pool *pool, enum ttm_caching caching,
 			       unsigned int order, struct page *p, bool reclaim)
@@ -223,9 +231,7 @@ static void ttm_pool_free_page(struct ttm_pool *pool, enum ttm_caching caching,
 #endif
 
 	if (!pool || !ttm_pool_uses_dma_alloc(pool)) {
-		mod_lruvec_page_state(p, reclaim ? NR_GPU_RECLAIM : NR_GPU_ACTIVE,
-				      -(1 << order));
-		__free_pages(p, order);
+		__free_pages_gpu_account(p, order, reclaim);
 		return;
 	}
 
@@ -606,7 +612,7 @@ static int ttm_pool_restore_commit(struct ttm_pool_tt_restore *restore,
 			 */
 			ttm_pool_split_for_swap(restore->pool, p);
 			copy_highpage(restore->alloced_page + i, p);
-			__free_pages(p, 0);
+			__free_pages_gpu_account(p, 0, false);
 		}
 
 		restore->restored_pages++;
@@ -1068,7 +1074,7 @@ long ttm_pool_backup(struct ttm_pool *pool, struct ttm_tt *tt,
 			if (flags->purge) {
 				shrunken += num_pages;
 				page->private = 0;
-				__free_pages(page, order);
+				__free_pages_gpu_account(page, order, false);
 				memset(tt->pages + i, 0,
 				       num_pages * sizeof(*tt->pages));
 			}
-- 
2.34.1

Re: [PATCH] drm/ttm: Fix GPU MM stats during pool shrinking

[Matthew Brost]

On Fri, May 01, 2026 at 05:05:14PM -0700, Kenneth Crudup wrote:
> 
> On 5/1/26 15:30, Matthew Brost wrote:
> 
> > TTM pool shrinking frees pages by calling __free_pages() directly,
> > which bypasses updates to NR_GPU_ACTIVE and leaves GPU MM accounting
> > out of sync.
> > 
> > Introduce a helper, __free_pages_gpu_account(), and use it for all page
> > frees in ttm_pool.c so GPU MM statistics are updated consistently.
> 
> OK, so why/how does "bonnie++" increase the GPU Memory size?
> 

Well, it shouldn’t. What bonnie++ does is basically consume system
memory, triggering reclaim, which in turn will evict GPU BOs that exist
when the display is open. Thanks for pointing me to bonnie++ - this,
plus running something like the WebGL Aquarium in Chrome, is a very nice
test case for Xe/DRM shrinkers.

> ----
> SwapTotal:      33554428 kB
> MemTotal:       32345672 kB
> GPUReclaim:       621568 kB
> GPUActive:        453592 kB
> 
> SwapTotal:      33554428 kB
> MemTotal:       32345672 kB
> GPUActive:       5554976 kB
> GPUReclaim:        12716 kB
> 
> SwapTotal:      33554428 kB
> MemTotal:       32345672 kB
> GPUActive:      18407272 kB
> GPUReclaim:          884 kB
> 
> SwapTotal:      33554428 kB
> MemTotal:       32345672 kB
> GPUActive:      24022916 kB
> GPUReclaim:          716 kB
> 
> SwapTotal:      33554428 kB
> MemTotal:       32345672 kB
> GPUActive:      25258248 kB
> GPUReclaim:        16032 kB
> 
> SwapTotal:      33554428 kB
> MemTotal:       32345672 kB
> GPUActive:      28207188 kB
> GPUReclaim:         3684 kB
> ----
> 
> ... and I'm now not so sure the patch is working ... this after a 2nd bonnie
> run:

It doesn’t appear that it is. I can recreate what you’re seeing with
this patch alone on drm-tip. I originally coded this patch on top of a
local fix to avoid the TTM shrinker allocating higher-order folios when
reclaiming memory—this is working there. I falsely assumed it would work
on drm-tip as well. Let me ensure I have a standalone fix for GPUActive
accounting first, then apply my TTM shrinker fix on top.

Matt

> 
> ----
> GPUActive:      44357100 kB
> SwapTotal:      33554428 kB
> MemTotal:       32345672 kB
> GPUReclaim:        94864 kB
> 
> GPUActive:      44373904 kB
> SwapTotal:      33554428 kB
> MemTotal:       32345672 kB
> GPUReclaim:        94996 kB
> 
> GPUActive:      44354940 kB
> SwapTotal:      33554428 kB
> MemTotal:       32345672 kB
> GPUReclaim:        98048 kB
> 
> GPUActive:      44769340 kB
> SwapTotal:      33554428 kB
> MemTotal:       32345672 kB
> GPUReclaim:       122996 kB
> ----
> 
> -Kenny
> 
> -- 
> Kenneth R. Crudup / Sr. SW Engineer, Scott County Consulting, Orange County
> CA
> 

Re: [PATCH] drm/ttm: Fix GPU MM stats during pool shrinking

[Matthew Brost]

On Fri, May 01, 2026 at 09:12:31PM -0700, Matthew Brost wrote:
> On Fri, May 01, 2026 at 05:05:14PM -0700, Kenneth Crudup wrote:
> > 
> > On 5/1/26 15:30, Matthew Brost wrote:
> > 
> > > TTM pool shrinking frees pages by calling __free_pages() directly,
> > > which bypasses updates to NR_GPU_ACTIVE and leaves GPU MM accounting
> > > out of sync.
> > > 
> > > Introduce a helper, __free_pages_gpu_account(), and use it for all page
> > > frees in ttm_pool.c so GPU MM statistics are updated consistently.
> > 
> > OK, so why/how does "bonnie++" increase the GPU Memory size?
> > 
> 
> Well, it shouldn’t. What bonnie++ does is basically consume system
> memory, triggering reclaim, which in turn will evict GPU BOs that exist
> when the display is open. Thanks for pointing me to bonnie++ - this,
> plus running something like the WebGL Aquarium in Chrome, is a very nice
> test case for Xe/DRM shrinkers.
> 
> > ----
> > SwapTotal:      33554428 kB
> > MemTotal:       32345672 kB
> > GPUReclaim:       621568 kB
> > GPUActive:        453592 kB
> > 
> > SwapTotal:      33554428 kB
> > MemTotal:       32345672 kB
> > GPUActive:       5554976 kB
> > GPUReclaim:        12716 kB
> > 
> > SwapTotal:      33554428 kB
> > MemTotal:       32345672 kB
> > GPUActive:      18407272 kB
> > GPUReclaim:          884 kB
> > 
> > SwapTotal:      33554428 kB
> > MemTotal:       32345672 kB
> > GPUActive:      24022916 kB
> > GPUReclaim:          716 kB
> > 
> > SwapTotal:      33554428 kB
> > MemTotal:       32345672 kB
> > GPUActive:      25258248 kB
> > GPUReclaim:        16032 kB
> > 
> > SwapTotal:      33554428 kB
> > MemTotal:       32345672 kB
> > GPUActive:      28207188 kB
> > GPUReclaim:         3684 kB
> > ----
> > 
> > ... and I'm now not so sure the patch is working ... this after a 2nd bonnie
> > run:
> 
> It doesn’t appear that it is. I can recreate what you’re seeing with
> this patch alone on drm-tip. I originally coded this patch on top of a
> local fix to avoid the TTM shrinker allocating higher-order folios when

s/allocating/spliting

Matt

> reclaiming memory—this is working there. I falsely assumed it would work
> on drm-tip as well. Let me ensure I have a standalone fix for GPUActive
> accounting first, then apply my TTM shrinker fix on top.
> 
> Matt
> 
> > 
> > ----
> > GPUActive:      44357100 kB
> > SwapTotal:      33554428 kB
> > MemTotal:       32345672 kB
> > GPUReclaim:        94864 kB
> > 
> > GPUActive:      44373904 kB
> > SwapTotal:      33554428 kB
> > MemTotal:       32345672 kB
> > GPUReclaim:        94996 kB
> > 
> > GPUActive:      44354940 kB
> > SwapTotal:      33554428 kB
> > MemTotal:       32345672 kB
> > GPUReclaim:        98048 kB
> > 
> > GPUActive:      44769340 kB
> > SwapTotal:      33554428 kB
> > MemTotal:       32345672 kB
> > GPUReclaim:       122996 kB
> > ----
> > 
> > -Kenny
> > 
> > -- 
> > Kenneth R. Crudup / Sr. SW Engineer, Scott County Consulting, Orange County
> > CA
> > 

Claude review: drm/ttm: Fix GPU MM stats during pool shrinking

[Claude Code Review Bot]

Overall Series Review

Subject: drm/ttm: Fix GPU MM stats during pool shrinking
Author: Matthew Brost <matthew.brost@intel.com>
Patches: 3
Reviewed: 2026-05-05T09:02:45.099837

---

This is a single-patch fix for a GPU memory accounting bug in TTM pool management. The issue is real and the fix is correct: two call sites in `ttm_pool.c` were calling `__free_pages()` directly without decrementing `NR_GPU_ACTIVE`, causing GPU MM stats to drift upward over time as pages are freed through the backup/restore and purge paths. The approach of extracting a helper (`__free_pages_gpu_account()`) is clean and ensures consistency across all three `__free_pages` call sites in the file.

The patch is minimal, well-scoped, and the Fixes tag correctly identifies the commit that introduced the GPU MM accounting.

**Recommendation: Approve.** One minor question worth confirming below, but it shouldn't block the patch.

---

---
Generated by Claude Code Patch Reviewer

Claude review: drm/ttm: Fix GPU MM stats during pool shrinking

[Claude Code Review Bot]

Patch Review

**Correctness: Good.**

The patch introduces:
```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);
}
```

And applies it to all three `__free_pages` call sites:

1. **`ttm_pool_free_page()` (line 225)** -- Pure refactor, replaces identical inline code. Already guarded by `if (!pool || !ttm_pool_uses_dma_alloc(pool))`, so it only runs for non-DMA pages that were accounted at allocation. Correct.

2. **`ttm_pool_restore_commit()` (line 609)** -- This frees an old retained page after splitting and copying to a new allocation. These pages were previously allocated through `ttm_pool_alloc_page()` which accounts them as `NR_GPU_ACTIVE` (line 175 in the current tree). Passing `reclaim=false` correctly decrements `NR_GPU_ACTIVE`. This was a real leak.

3. **`ttm_pool_backup()` (line 1071)** -- The purge path frees pages directly from `tt->pages`. These are actively-used pages counted in `NR_GPU_ACTIVE` (either from initial allocation or from `ttm_pool_type_take()` which moves pages back to ACTIVE state). Passing `reclaim=false` is correct. This was the main leak reported by the user.

**Coverage check:** I verified that these are the only three `__free_pages` call sites in `ttm_pool.c`:
- Line 228: handled (refactored)
- Line 609: handled (bug fix)
- Line 1071: handled (bug fix)

The shrinker dispose path (`ttm_pool_dispose_list` -> `ttm_pool_free_page` with `reclaim=true`) was already correct since it went through `ttm_pool_free_page()`.

**One question worth confirming:** At lines 609 and 1071, the `__free_pages()` call is not guarded by a `!ttm_pool_uses_dma_alloc()` check (unlike `ttm_pool_free_page` which has that guard). If pages from a DMA-alloc pool could reach these paths, calling `__free_pages_gpu_account()` would undercount `NR_GPU_ACTIVE` (since DMA-alloc pages never increment it at allocation). However, this would also mean the pre-existing `__free_pages()` call was itself wrong for DMA pages (leaking the DMA allocation), so this patch doesn't make it worse -- it's a separate pre-existing concern. It's worth confirming that these paths are indeed only reachable for non-DMA pools, or if not, whether there's a separate fix needed for the DMA-alloc case.

---
Generated by Claude Code Patch Reviewer