drm/amdgpu: only support IBs in the buffer list (v2)

amdgpu_cs_find_mapping doesn't work without all buffers being validated,
so the TTM validation must be done first.

v2: only use amdgpu_cs_find_mapping for UVD/VCE VM emulation
Signed-off-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>

drivers/gpu/drm/amd/amdgpu/amdgpu.h

@@ -1191,7 +1191,6 @@ struct amdgpu_cs_parser {
 	struct amdgpu_cs_chunk	*chunks;
 	/* relocations */
 	struct amdgpu_bo_list_entry	*vm_bos;
-	struct amdgpu_bo_list_entry	*ib_bos;
 	struct list_head	validated;
 
 	struct amdgpu_ib	*ibs;

drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c

@@ -230,11 +230,6 @@ int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
 		goto out;
 	}
 
-	p->ib_bos = kcalloc(p->num_ibs, sizeof(struct amdgpu_bo_list_entry),
-			    GFP_KERNEL);
-	if (!p->ib_bos)
-		r = -ENOMEM;
-
 out:
 	kfree(chunk_array);
 	return r;
@@ -373,13 +368,6 @@ static int amdgpu_cs_parser_relocs(struct amdgpu_cs_parser *p)
 	p->vm_bos = amdgpu_vm_get_bos(p->adev, &fpriv->vm,
 				      &p->validated);
 
-	for (i = 0; i < p->num_ibs; i++) {
-		if (!p->ib_bos[i].robj)
-			continue;
-
-		list_add(&p->ib_bos[i].tv.head, &p->validated);
-	}
-
 	if (need_mmap_lock)
 		down_read(&current->mm->mmap_sem);
 
@@ -457,15 +445,9 @@ static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error, bo
 	for (i = 0; i < parser->nchunks; i++)
 		drm_free_large(parser->chunks[i].kdata);
 	kfree(parser->chunks);
-	for (i = 0; i < parser->num_ibs; i++) {
-		struct amdgpu_bo *bo = parser->ib_bos[i].robj;
+	for (i = 0; i < parser->num_ibs; i++)
 		amdgpu_ib_free(parser->adev, &parser->ibs[i]);
-
-		if (bo)
-			drm_gem_object_unreference_unlocked(&bo->gem_base);
-	}
 	kfree(parser->ibs);
-	kfree(parser->ib_bos);
 	if (parser->uf.bo)
 		drm_gem_object_unreference_unlocked(&parser->uf.bo->gem_base);
 }
@@ -505,21 +487,6 @@ static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p,
 		}
 	}
 
-	for (i = 0; i < p->num_ibs; i++) {
-		bo = p->ib_bos[i].robj;
-		if (!bo)
-			continue;
-
-		bo_va = p->ib_bos[i].bo_va;
-		if (!bo_va)
-			continue;
-
-		r = amdgpu_vm_bo_update(adev, bo_va, &bo->tbo.mem);
-		if (r)
-			return r;
-
-		amdgpu_sync_fence(&p->ibs[0].sync, bo_va->last_pt_update);
-	}
 	return amdgpu_vm_clear_invalids(adev, vm, &p->ibs[0].sync);
 }
 
@@ -581,11 +548,7 @@ static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
 		struct amdgpu_cs_chunk *chunk;
 		struct amdgpu_ib *ib;
 		struct drm_amdgpu_cs_chunk_ib *chunk_ib;
-		struct amdgpu_bo_list_entry *ib_bo;
 		struct amdgpu_ring *ring;
-		struct drm_gem_object *gobj;
-		struct amdgpu_bo *aobj;
-		void *kptr;
 
 		chunk = &parser->chunks[i];
 		ib = &parser->ibs[j];
@@ -594,66 +557,49 @@ static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
 		if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
 			continue;
 
-		gobj = drm_gem_object_lookup(adev->ddev, parser->filp, chunk_ib->handle);
-		if (gobj == NULL)
-			return -ENOENT;
-		aobj = gem_to_amdgpu_bo(gobj);
-
 		r = amdgpu_cs_get_ring(adev, chunk_ib->ip_type,
 				       chunk_ib->ip_instance, chunk_ib->ring,
 				       &ring);
-		if (r) {
-			drm_gem_object_unreference_unlocked(gobj);
+		if (r)
 			return r;
-		}
 
 		if (ring->funcs->parse_cs) {
-			r = amdgpu_bo_reserve(aobj, false);
-			if (r) {
-				drm_gem_object_unreference_unlocked(gobj);
-				return r;
+			struct amdgpu_bo *aobj = NULL;
+			void *kptr;
+
+			amdgpu_cs_find_mapping(parser, chunk_ib->va_start, &aobj);
+			if (!aobj) {
+				DRM_ERROR("IB va_start is invalid\n");
+				return -EINVAL;
 			}
 
+			/* the IB should be reserved at this point */
 			r = amdgpu_bo_kmap(aobj, &kptr);
 			if (r) {
-				amdgpu_bo_unreserve(aobj);
-				drm_gem_object_unreference_unlocked(gobj);
 				return r;
 			}
 
 			r =  amdgpu_ib_get(ring, NULL, chunk_ib->ib_bytes, ib);
 			if (r) {
 				DRM_ERROR("Failed to get ib !\n");
-				amdgpu_bo_unreserve(aobj);
-				drm_gem_object_unreference_unlocked(gobj);
 				return r;
 			}
 
 			memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
 			amdgpu_bo_kunmap(aobj);
-			amdgpu_bo_unreserve(aobj);
 		} else {
 			r =  amdgpu_ib_get(ring, vm, 0, ib);
 			if (r) {
 				DRM_ERROR("Failed to get ib !\n");
-				drm_gem_object_unreference_unlocked(gobj);
 				return r;
 			}
 
 			ib->gpu_addr = chunk_ib->va_start;
 		}
-		ib->length_dw = chunk_ib->ib_bytes / 4;
 
+		ib->length_dw = chunk_ib->ib_bytes / 4;
 		ib->flags = chunk_ib->flags;
 		ib->ctx = parser->ctx;
-
-		ib_bo = &parser->ib_bos[j];
-		ib_bo->robj = aobj;
-		ib_bo->prefered_domains = aobj->initial_domain;
-		ib_bo->allowed_domains = aobj->initial_domain;
-		ib_bo->priority = 0;
-		ib_bo->tv.bo = &aobj->tbo;
-		ib_bo->tv.shared = true;
 		j++;
 	}
 
@@ -702,6 +648,7 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 	union drm_amdgpu_cs *cs = data;
 	struct amdgpu_cs_parser parser;
 	int r, i;
+	bool reserved_buffers = false;
 
 	down_read(&adev->exclusive_lock);
 	if (!adev->accel_working) {
@@ -721,15 +668,21 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 		return r;
 	}
 
-	r = amdgpu_cs_ib_fill(adev, &parser);
-	if (!r) {
 	r = amdgpu_cs_parser_relocs(&parser);
-		if (r && r != -ERESTARTSYS)
-			DRM_ERROR("Failed to parse relocation %d!\n", r);
+	if (r) {
+		if (r != -ERESTARTSYS) {
+			if (r == -ENOMEM)
+				DRM_ERROR("Not enough memory for command submission!\n");
+			else
+				DRM_ERROR("Failed to process the buffer list %d!\n", r);
+		}
+	} else {
+		reserved_buffers = true;
+		r = amdgpu_cs_ib_fill(adev, &parser);
 	}
 
 	if (r) {
-		amdgpu_cs_parser_fini(&parser, r, false);
+		amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
 		up_read(&adev->exclusive_lock);
 		r = amdgpu_cs_handle_lockup(adev, r);
 		return r;

include/uapi/drm/amdgpu_drm.h

@@ -378,11 +378,7 @@ union drm_amdgpu_cs {
 #define AMDGPU_IB_FLAG_PREAMBLE (1<<2)
 
 struct drm_amdgpu_cs_chunk_ib {
-	/**
-	 * Handle of GEM object to be used as IB or 0 if it is already in
-	 * residency list.
-	 */
-	uint32_t handle;
+	uint32_t _pad;
 	uint32_t flags;		/* IB Flags */
 	uint64_t va_start;	/* Virtual address to begin IB execution */
 	uint32_t ib_bytes;	/* Size of submission */