From: Lucas Frendorf <[email protected]>
ttm_pool_alloc_page() and ttm_pool_free_page() account NR_GPU_ACTIVE
and NR_GPU_RECLAIM only for pages from the page allocator, but
ttm_pool_type_give() and ttm_pool_type_take() update them for every
page. Each DMA-allocated page that is given to a pool and later freed
from it therefore leaves NR_GPU_RECLAIM incremented and NR_GPU_ACTIVE
decremented.
DMA-allocating pools are used by amdgpu, radeon and nouveau when
drm_need_swiotlb() is true, and by vmwgfx in coherent mode. On an
amdgpu Rembrandt APU, GPUReclaim reached 20 GiB on a 30.6 GiB machine
after 2 days while the pool held 34 MiB, and GPUActive read 0 kB.
The imbalance was found with LLM assistance while investigating a
memory monitor that reported 95% use.
The allocation and free paths do not account DMA API memory. Skip the
give and take accounting for DMA-allocating pools as well.
Fixes: ae80122f3896 ("drm/ttm: use gpu mm stats to track gpu memory
allocations. (v4)")
Link: https://github.com/AvengeMedia/dgop/issues/44
Cc: Christian Koenig <[email protected]>
Cc: Huang Rui <[email protected]>
Cc: Matthew Auld <[email protected]>
Cc: Matthew Brost <[email protected]>
Cc: David Airlie <[email protected]>
Cc: [email protected]
Assisted-by: LLM
Signed-off-by: Lucas Frendorf <[email protected]>
---
I noticed dgop, a memory monitor, reported 95% usage. An LLM found the
cause in the GPUReclaim accounting, wrote this patch and drafted its
changelog, which I reviewed and revised.
To reproduce, compare GPUReclaim in /proc/meminfo with the page total
in /sys/kernel/debug/dri/<pci-id>/ttm_page_pool on a device with a
DMA-allocating pool.
Tested on v7.2.9 with the patched ttm.ko: GPUReclaim stayed at 0 kB
through a glmark2 run on the 680M while its DMA pool held 298 MiB. The
RX 6800S, which uses the non-DMA pools, still raised and lowered
GPUReclaim normally. The patch builds with W=1 on v7.3-rc6.
Dave Airlie's memcg series (part 2 v2, patch 05/10) rewrites both
hunks and would need the same check.
---
drivers/gpu/drm/ttm/ttm_pool.c | 12 ++++++++----
1 file changed, 8 insertions(+), 4 deletions(-)
diff --git a/drivers/gpu/drm/ttm/ttm_pool.c b/drivers/gpu/drm/ttm/ttm_pool.c
index 1bf37023f..1f2d31c3e 100644
--- a/drivers/gpu/drm/ttm/ttm_pool.c
+++ b/drivers/gpu/drm/ttm/ttm_pool.c
@@ -341,8 +341,10 @@ static void ttm_pool_type_give(struct ttm_pool_type *pt,
struct page *p)
rcu_read_unlock();
atomic_long_add(num_pages, &allocated_pages[nid]);
- mod_lruvec_page_state(p, NR_GPU_ACTIVE, -num_pages);
- mod_lruvec_page_state(p, NR_GPU_RECLAIM, num_pages);
+ if (!pt->pool || !ttm_pool_uses_dma_alloc(pt->pool)) {
+ mod_lruvec_page_state(p, NR_GPU_ACTIVE, -num_pages);
+ mod_lruvec_page_state(p, NR_GPU_RECLAIM, num_pages);
+ }
}
static enum lru_status take_one_from_lru(struct list_head *item,
@@ -367,8 +369,10 @@ static struct page *ttm_pool_type_take(struct
ttm_pool_type *pt, int nid)
ret = list_lru_walk_node(&pt->pages, nid, take_one_from_lru, (void
*)&p, &nr_to_walk);
if (ret == 1 && p) {
atomic_long_sub(1 << pt->order, &allocated_pages[nid]);
- mod_lruvec_page_state(p, NR_GPU_ACTIVE, (1 << pt->order));
- mod_lruvec_page_state(p, NR_GPU_RECLAIM, -(1 << pt->order));
+ if (!pt->pool || !ttm_pool_uses_dma_alloc(pt->pool)) {
+ mod_lruvec_page_state(p, NR_GPU_ACTIVE, (1 <<
pt->order));
+ mod_lruvec_page_state(p, NR_GPU_RECLAIM, -(1 <<
pt->order));
+ }
}
return p;
}
---
base-commit: a90ee4305c4a5df72c11b31dacfdc76e00fcf78a
change-id: 20261008-ttm-gpu-stats-dma-9fc5e952a45b
Best regards,
--
Lucas Frendorf <[email protected]>
--
Lucas Frendorf <[email protected]>
