1 files changed, 66 insertions, 30 deletions

diff --git a/drivers/infiniband/hw/mlx5/umr.c b/drivers/infiniband/hw/mlx5/umr.c
index ffc31b01f690..887fd6fa3ba9 100644
--- a/drivers/infiniband/hw/mlx5/umr.c
+++ b/drivers/infiniband/hw/mlx5/umr.c
@@ -224,6 +224,9 @@ int mlx5r_umr_init(struct mlx5_ib_dev *dev)
 
 void mlx5r_umr_cleanup(struct mlx5_ib_dev *dev)
 {
+	if (!dev->umrc.pd)
+		return;
+
 	mutex_destroy(&dev->umrc.init_lock);
 	ib_dealloc_pd(dev->umrc.pd);
 }
@@ -632,44 +635,47 @@ static void mlx5r_umr_final_update_xlt(struct mlx5_ib_dev *dev,
 	wqe->data_seg.byte_count = cpu_to_be32(sg->length);
 }
 
-/*
- * Send the DMA list to the HW for a normal MR using UMR.
- * Dmabuf MR is handled in a similar way, except that the MLX5_IB_UPD_XLT_ZAP
- * flag may be used.
- */
-int mlx5r_umr_update_mr_pas(struct mlx5_ib_mr *mr, unsigned int flags)
+static int
+_mlx5r_umr_update_mr_pas(struct mlx5_ib_mr *mr, unsigned int flags, bool dd)
 {
+	size_t ent_size = dd ? sizeof(struct mlx5_ksm) : sizeof(struct mlx5_mtt);
 	struct mlx5_ib_dev *dev = mr_to_mdev(mr);
 	struct device *ddev = &dev->mdev->pdev->dev;
 	struct mlx5r_umr_wqe wqe = {};
 	struct ib_block_iter biter;
+	struct mlx5_ksm *cur_ksm;
 	struct mlx5_mtt *cur_mtt;
 	size_t orig_sg_length;
-	struct mlx5_mtt *mtt;
 	size_t final_size;
+	void *curr_entry;
 	struct ib_sge sg;
+	void *entry;
 	u64 offset = 0;
 	int err = 0;
 
-	if (WARN_ON(mr->umem->is_odp))
-		return -EINVAL;
-
-	mtt = mlx5r_umr_create_xlt(
-		dev, &sg, ib_umem_num_dma_blocks(mr->umem, 1 << mr->page_shift),
-		sizeof(*mtt), flags);
-	if (!mtt)
+	entry = mlx5r_umr_create_xlt(dev, &sg,
+				     ib_umem_num_dma_blocks(mr->umem, 1 << mr->page_shift),
+				     ent_size, flags);
+	if (!entry)
 		return -ENOMEM;
 
 	orig_sg_length = sg.length;
-
 	mlx5r_umr_set_update_xlt_ctrl_seg(&wqe.ctrl_seg, flags, &sg);
 	mlx5r_umr_set_update_xlt_mkey_seg(dev, &wqe.mkey_seg, mr,
 					  mr->page_shift);
+	if (dd) {
+		/* Use the data direct internal kernel PD */
+		MLX5_SET(mkc, &wqe.mkey_seg, pd, dev->ddr.pdn);
+		cur_ksm = entry;
+	} else {
+		cur_mtt = entry;
+	}
+
 	mlx5r_umr_set_update_xlt_data_seg(&wqe.data_seg, &sg);
 
-	cur_mtt = mtt;
+	curr_entry = entry;
 	rdma_umem_for_each_dma_block(mr->umem, &biter, BIT(mr->page_shift)) {
-		if (cur_mtt == (void *)mtt + sg.length) {
+		if (curr_entry == entry + sg.length) {
 			dma_sync_single_for_device(ddev, sg.addr, sg.length,
 						   DMA_TO_DEVICE);
 
@@ -681,23 +687,31 @@ int mlx5r_umr_update_mr_pas(struct mlx5_ib_mr *mr, unsigned int flags)
 						DMA_TO_DEVICE);
 			offset += sg.length;
 			mlx5r_umr_update_offset(&wqe.ctrl_seg, offset);
-
-			cur_mtt = mtt;
+			if (dd)
+				cur_ksm = entry;
+			else
+				cur_mtt = entry;
 		}
 
-		cur_mtt->ptag =
-			cpu_to_be64(rdma_block_iter_dma_address(&biter) |
-				    MLX5_IB_MTT_PRESENT);
-
-		if (mr->umem->is_dmabuf && (flags & MLX5_IB_UPD_XLT_ZAP))
-			cur_mtt->ptag = 0;
-
-		cur_mtt++;
+		if (dd) {
+			cur_ksm->va = cpu_to_be64(rdma_block_iter_dma_address(&biter));
+			cur_ksm->key = cpu_to_be32(dev->ddr.mkey);
+			cur_ksm++;
+			curr_entry = cur_ksm;
+		} else {
+			cur_mtt->ptag =
+				cpu_to_be64(rdma_block_iter_dma_address(&biter) |
+					    MLX5_IB_MTT_PRESENT);
+			if (mr->umem->is_dmabuf && (flags & MLX5_IB_UPD_XLT_ZAP))
+				cur_mtt->ptag = 0;
+			cur_mtt++;
+			curr_entry = cur_mtt;
+		}
 	}
 
-	final_size = (void *)cur_mtt - (void *)mtt;
+	final_size = curr_entry - entry;
 	sg.length = ALIGN(final_size, MLX5_UMR_FLEX_ALIGNMENT);
-	memset(cur_mtt, 0, sg.length - final_size);
+	memset(curr_entry, 0, sg.length - final_size);
 	mlx5r_umr_final_update_xlt(dev, &wqe, mr, &sg, flags);
 
 	dma_sync_single_for_device(ddev, sg.addr, sg.length, DMA_TO_DEVICE);
@@ -705,10 +719,32 @@ int mlx5r_umr_update_mr_pas(struct mlx5_ib_mr *mr, unsigned int flags)
 
 err:
 	sg.length = orig_sg_length;
-	mlx5r_umr_unmap_free_xlt(dev, mtt, &sg);
+	mlx5r_umr_unmap_free_xlt(dev, entry, &sg);
 	return err;
 }
 
+int mlx5r_umr_update_data_direct_ksm_pas(struct mlx5_ib_mr *mr, unsigned int flags)
+{
+	/* No invalidation flow is expected */
+	if (WARN_ON(!mr->umem->is_dmabuf) || (flags & MLX5_IB_UPD_XLT_ZAP))
+		return -EINVAL;
+
+	return _mlx5r_umr_update_mr_pas(mr, flags, true);
+}
+
+/*
+ * Send the DMA list to the HW for a normal MR using UMR.
+ * Dmabuf MR is handled in a similar way, except that the MLX5_IB_UPD_XLT_ZAP
+ * flag may be used.
+ */
+int mlx5r_umr_update_mr_pas(struct mlx5_ib_mr *mr, unsigned int flags)
+{
+	if (WARN_ON(mr->umem->is_odp))
+		return -EINVAL;
+
+	return _mlx5r_umr_update_mr_pas(mr, flags, false);
+}
+
 static bool umr_can_use_indirect_mkey(struct mlx5_ib_dev *dev)
 {
 	return !MLX5_CAP_GEN(dev->mdev, umr_indirect_mkey_disabled);