Try to allocate system memory on the NUMA node the device is closest to and try to run delayed_delete workers on a CPU of this node as well. To optimize the memory clearing operation when a TTM BO gets freed by the delayed_delete worker, scheduling it closer to a NUMA node where the memory was initially allocated helps avoid the cases where the worker gets randomly scheduled on the CPU cores that are across interconnect boundaries such as xGMI, PCIe etc. This change helps USWC GTT allocations on NUMA systems (dGPU) and AMD APU platforms such as GFXIP9.4.3. Acked-by: Felix Kuehling <Felix.Kuehling@xxxxxxx> Signed-off-by: Rajneesh Bhardwaj <rajneesh.bhardwaj@xxxxxxx> --- Changes in v2: - Absorbed the feedback provided by Christian in the commit message and the comment. drivers/gpu/drm/ttm/ttm_bo.c | 8 +++++++- drivers/gpu/drm/ttm/ttm_device.c | 3 ++- 2 files changed, 9 insertions(+), 2 deletions(-) diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c index 5757b9415e37..6f28a77a565b 100644 --- a/drivers/gpu/drm/ttm/ttm_bo.c +++ b/drivers/gpu/drm/ttm/ttm_bo.c @@ -370,7 +370,13 @@ static void ttm_bo_release(struct kref *kref) spin_unlock(&bo->bdev->lru_lock); INIT_WORK(&bo->delayed_delete, ttm_bo_delayed_delete); - queue_work(bdev->wq, &bo->delayed_delete); + + /* Schedule the worker on the closest NUMA node. This + * improves performance since system memory might be + * cleared on free and that is best done on a CPU core + * close to it. + */ + queue_work_node(bdev->pool.nid, bdev->wq, &bo->delayed_delete); return; } diff --git a/drivers/gpu/drm/ttm/ttm_device.c b/drivers/gpu/drm/ttm/ttm_device.c index 43e27ab77f95..72b81a2ee6c7 100644 --- a/drivers/gpu/drm/ttm/ttm_device.c +++ b/drivers/gpu/drm/ttm/ttm_device.c @@ -213,7 +213,8 @@ int ttm_device_init(struct ttm_device *bdev, struct ttm_device_funcs *funcs, bdev->funcs = funcs; ttm_sys_man_init(bdev); - ttm_pool_init(&bdev->pool, dev, NUMA_NO_NODE, use_dma_alloc, use_dma32); + + ttm_pool_init(&bdev->pool, dev, dev_to_node(dev), use_dma_alloc, use_dma32); bdev->vma_manager = vma_manager; spin_lock_init(&bdev->lru_lock); -- 2.34.1
