drm/tegra: gem: Don't attach dma-bufs when not needed

The dma-buf import code currently attaches and maps all imported
dma-bufs to the drm device to get their sgt for mapping to the
directly managed IOMMU domain.

In many cases, like for newer chips (Tegra186+), the directly
managed IOMMU domain is, however, not used. Mapping to the drm
device can also cause issues e.g. with swiotlb since it is not
a real device.

To improve the situation, only attach and map imported dma-bufs
when required.

Signed-off-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240424051335.2872574-2-cyndis@kapsi.fi

Author: cyndis

drivers/gpu/drm/tegra/gem.c

@@ -76,8 +76,8 @@ static struct host1x_bo_mapping *tegra_bo_pin(struct device *dev, struct host1x_
 	/*
 	 * Imported buffers need special treatment to satisfy the semantics of DMA-BUF.
 	 */
-	if (gem->import_attach) {
-		struct dma_buf *buf = gem->import_attach->dmabuf;
+	if (obj->dma_buf) {
+		struct dma_buf *buf = obj->dma_buf;
 
 		map->attach = dma_buf_attach(buf, dev);
 		if (IS_ERR(map->attach)) {
@@ -184,8 +184,8 @@ static void *tegra_bo_mmap(struct host1x_bo *bo)
 	if (obj->vaddr)
 		return obj->vaddr;
 
-	if (obj->gem.import_attach) {
-		ret = dma_buf_vmap_unlocked(obj->gem.import_attach->dmabuf, &map);
+	if (obj->dma_buf) {
+		ret = dma_buf_vmap_unlocked(obj->dma_buf, &map);
 		if (ret < 0)
 			return ERR_PTR(ret);
 
@@ -208,8 +208,8 @@ static void tegra_bo_munmap(struct host1x_bo *bo, void *addr)
 	if (obj->vaddr)
 		return;
 
-	if (obj->gem.import_attach)
-		return dma_buf_vunmap_unlocked(obj->gem.import_attach->dmabuf, &map);
+	if (obj->dma_buf)
+		return dma_buf_vunmap_unlocked(obj->dma_buf, &map);
 
 	vunmap(addr);
 }
@@ -465,27 +465,32 @@ static struct tegra_bo *tegra_bo_import(struct drm_device *drm,
 	if (IS_ERR(bo))
 		return bo;
 
-	attach = dma_buf_attach(buf, drm->dev);
-	if (IS_ERR(attach)) {
-		err = PTR_ERR(attach);
-		goto free;
-	}
-
-	get_dma_buf(buf);
+	/*
+	 * If we need to use IOMMU API to map the dma-buf into the internally managed
+	 * domain, map it first to the DRM device to get an sgt.
+	 */
+	if (tegra->domain) {
+		attach = dma_buf_attach(buf, drm->dev);
+		if (IS_ERR(attach)) {
+			err = PTR_ERR(attach);
+			goto free;
+		}
 
-	bo->sgt = dma_buf_map_attachment_unlocked(attach, DMA_TO_DEVICE);
-	if (IS_ERR(bo->sgt)) {
-		err = PTR_ERR(bo->sgt);
-		goto detach;
-	}
+		bo->sgt = dma_buf_map_attachment_unlocked(attach, DMA_TO_DEVICE);
+		if (IS_ERR(bo->sgt)) {
+			err = PTR_ERR(bo->sgt);
+			goto detach;
+		}
 
-	if (tegra->domain) {
 		err = tegra_bo_iommu_map(tegra, bo);
 		if (err < 0)
 			goto detach;
+
+		bo->gem.import_attach = attach;
 	}
 
-	bo->gem.import_attach = attach;
+	get_dma_buf(buf);
+	bo->dma_buf = buf;
 
 	return bo;
 
@@ -516,20 +521,21 @@ void tegra_bo_free_object(struct drm_gem_object *gem)
 				dev_name(mapping->dev));
 	}
 
-	if (tegra->domain)
+	if (tegra->domain) {
 		tegra_bo_iommu_unmap(tegra, bo);
 
-	if (gem->import_attach) {
-		struct dma_buf *dmabuf = gem->import_attach->dmabuf;
-
-		dma_buf_unmap_attachment_unlocked(gem->import_attach, bo->sgt,
-						  DMA_TO_DEVICE);
-		dma_buf_detach(dmabuf, gem->import_attach);
-		dma_buf_put(dmabuf);
-	} else {
-		tegra_bo_free(gem->dev, bo);
+		if (gem->import_attach) {
+			dma_buf_unmap_attachment_unlocked(gem->import_attach, bo->sgt,
+							  DMA_TO_DEVICE);
+			dma_buf_detach(gem->import_attach->dmabuf, gem->import_attach);
+		}
 	}
 
+	tegra_bo_free(gem->dev, bo);
+
+	if (bo->dma_buf)
+		dma_buf_put(bo->dma_buf);
+
 	drm_gem_object_release(gem);
 	kfree(bo);
 }

drivers/gpu/drm/tegra/gem.h

@@ -32,13 +32,34 @@ struct tegra_bo_tiling {
 	enum tegra_bo_sector_layout sector_layout;
 };
 
+/*
+ * How memory is referenced within a tegra_bo:
+ *
+ * Buffer source  | Mapping API(*)  | Fields
+ * ---------------+-----------------+---------------
+ * Allocated here | DMA API         | iova (IOVA mapped to drm->dev), vaddr (CPU VA)
+ *
+ * Allocated here | IOMMU API       | pages/num_pages (Phys. memory), sgt (Mapped to drm->dev),
+ *                                  | iova/size (Mapped to domain)
+ *
+ * Imported       | DMA API         | dma_buf (Imported dma_buf)
+ *
+ * Imported       | IOMMU API       | dma_buf (Imported dma_buf),
+ *                                  | gem->import_attach (Attachment on drm->dev),
+ *                                  | sgt (Mapped to drm->dev)
+ *                                  | iova/size (Mapped to domain)
+ *
+ * (*) If tegra->domain is set, i.e. TegraDRM IOMMU domain is directly managed through IOMMU API,
+ *     this is IOMMU API. Otherwise DMA API.
+ */
 struct tegra_bo {
 	struct drm_gem_object gem;
 	struct host1x_bo base;
 	unsigned long flags;
 	struct sg_table *sgt;
 	dma_addr_t iova;
 	void *vaddr;
+	struct dma_buf *dma_buf;
 
 	struct drm_mm_node *mm;
 	unsigned long num_pages;