RDMA/mlx5: Enable ATS support for MRs and umems

For mlx5 if ATS is enabled in the PCI config then the device will use ATS
requests for only certain DMA operations. This has to be opted in by the
SW side based on the mkey or umem settings.

ATS slows down the PCI performance, so it should only be set in cases when
it is needed. All of these cases revolve around optimizing PCI P2P
transfers and avoiding bad cases where the bus just doesn't work.

Link: https://lore.kernel.org/r/4-v1-bd147097458e+ede-umem_dmabuf_jgg@nvidia.comSigned-off-by: Jason Gunthorpe <jgg@nvidia.com>

drivers/infiniband/hw/mlx5/devx.c

@@ -2160,26 +2160,17 @@ static int UVERBS_HANDLER(MLX5_IB_METHOD_DEVX_SUBSCRIBE_EVENT)(
 
 static int devx_umem_get(struct mlx5_ib_dev *dev, struct ib_ucontext *ucontext,
 			 struct uverbs_attr_bundle *attrs,
-			 struct devx_umem *obj)
+			 struct devx_umem *obj, u32 access_flags)
 {
 	u64 addr;
 	size_t size;
-	u32 access;
 	int err;
 
 	if (uverbs_copy_from(&addr, attrs, MLX5_IB_ATTR_DEVX_UMEM_REG_ADDR) ||
 	    uverbs_copy_from(&size, attrs, MLX5_IB_ATTR_DEVX_UMEM_REG_LEN))
 		return -EFAULT;
 
-	err = uverbs_get_flags32(&access, attrs,
-				 MLX5_IB_ATTR_DEVX_UMEM_REG_ACCESS,
-				 IB_ACCESS_LOCAL_WRITE |
-				 IB_ACCESS_REMOTE_WRITE |
-				 IB_ACCESS_REMOTE_READ);
-	if (err)
-		return err;
-
-	err = ib_check_mr_access(&dev->ib_dev, access);
+	err = ib_check_mr_access(&dev->ib_dev, access_flags);
 	if (err)
 		return err;
 
@@ -2193,12 +2184,12 @@ static int devx_umem_get(struct mlx5_ib_dev *dev, struct ib_ucontext *ucontext,
 			return -EFAULT;
 
 		umem_dmabuf = ib_umem_dmabuf_get_pinned(
-			&dev->ib_dev, addr, size, dmabuf_fd, access);
+			&dev->ib_dev, addr, size, dmabuf_fd, access_flags);
 		if (IS_ERR(umem_dmabuf))
 			return PTR_ERR(umem_dmabuf);
 		obj->umem = &umem_dmabuf->umem;
 	} else {
-		obj->umem = ib_umem_get(&dev->ib_dev, addr, size, access);
+		obj->umem = ib_umem_get(&dev->ib_dev, addr, size, access_flags);
 		if (IS_ERR(obj->umem))
 			return PTR_ERR(obj->umem);
 	}
@@ -2240,7 +2231,8 @@ static unsigned int devx_umem_find_best_pgsize(struct ib_umem *umem,
 static int devx_umem_reg_cmd_alloc(struct mlx5_ib_dev *dev,
 				   struct uverbs_attr_bundle *attrs,
 				   struct devx_umem *obj,
-				   struct devx_umem_reg_cmd *cmd)
+				   struct devx_umem_reg_cmd *cmd,
+				   int access)
 {
 	unsigned long pgsz_bitmap;
 	unsigned int page_size;
@@ -2289,6 +2281,9 @@ static int devx_umem_reg_cmd_alloc(struct mlx5_ib_dev *dev,
 	MLX5_SET(umem, umem, page_offset,
 		 ib_umem_dma_offset(obj->umem, page_size));
 
+	if (mlx5_umem_needs_ats(dev, obj->umem, access))
+		MLX5_SET(umem, umem, ats, 1);
+
 	mlx5_ib_populate_pas(obj->umem, page_size, mtt,
 			     (obj->umem->writable ? MLX5_IB_MTT_WRITE : 0) |
 				     MLX5_IB_MTT_READ);
@@ -2306,20 +2301,30 @@ static int UVERBS_HANDLER(MLX5_IB_METHOD_DEVX_UMEM_REG)(
 	struct mlx5_ib_ucontext *c = rdma_udata_to_drv_context(
 		&attrs->driver_udata, struct mlx5_ib_ucontext, ibucontext);
 	struct mlx5_ib_dev *dev = to_mdev(c->ibucontext.device);
+	int access_flags;
 	int err;
 
 	if (!c->devx_uid)
 		return -EINVAL;
 
+	err = uverbs_get_flags32(&access_flags, attrs,
+				 MLX5_IB_ATTR_DEVX_UMEM_REG_ACCESS,
+				 IB_ACCESS_LOCAL_WRITE |
+				 IB_ACCESS_REMOTE_WRITE |
+				 IB_ACCESS_REMOTE_READ |
+				 IB_ACCESS_RELAXED_ORDERING);
+	if (err)
+		return err;
+
 	obj = kzalloc(sizeof(struct devx_umem), GFP_KERNEL);
 	if (!obj)
 		return -ENOMEM;
 
-	err = devx_umem_get(dev, &c->ibucontext, attrs, obj);
+	err = devx_umem_get(dev, &c->ibucontext, attrs, obj, access_flags);
 	if (err)
 		goto err_obj_free;
 
-	err = devx_umem_reg_cmd_alloc(dev, attrs, obj, &cmd);
+	err = devx_umem_reg_cmd_alloc(dev, attrs, obj, &cmd, access_flags);
 	if (err)
 		goto err_umem_release;
 

drivers/infiniband/hw/mlx5/mlx5_ib.h

@@ -1550,4 +1550,40 @@ static inline bool rt_supported(int ts_cap)
 	return ts_cap == MLX5_TIMESTAMP_FORMAT_CAP_REAL_TIME ||
 	       ts_cap == MLX5_TIMESTAMP_FORMAT_CAP_FREE_RUNNING_AND_REAL_TIME;
 }
+
+/*
+ * PCI Peer to Peer is a trainwreck. If no switch is present then things
+ * sometimes work, depending on the pci_distance_p2p logic for excluding broken
+ * root complexes. However if a switch is present in the path, then things get
+ * really ugly depending on how the switch is setup. This table assumes that the
+ * root complex is strict and is validating that all req/reps are matches
+ * perfectly - so any scenario where it sees only half the transaction is a
+ * failure.
+ *
+ * CR/RR/DT  ATS RO P2P
+ * 00X       X   X  OK
+ * 010       X   X  fails (request is routed to root but root never sees comp)
+ * 011       0   X  fails (request is routed to root but root never sees comp)
+ * 011       1   X  OK
+ * 10X       X   1  OK
+ * 101       X   0  fails (completion is routed to root but root didn't see req)
+ * 110       X   0  SLOW
+ * 111       0   0  SLOW
+ * 111       1   0  fails (completion is routed to root but root didn't see req)
+ * 111       1   1  OK
+ *
+ * Unfortunately we cannot reliably know if a switch is present or what the
+ * CR/RR/DT ACS settings are, as in a VM that is all hidden. Assume that
+ * CR/RR/DT is 111 if the ATS cap is enabled and follow the last three rows.
+ *
+ * For now assume if the umem is a dma_buf then it is P2P.
+ */
+static inline bool mlx5_umem_needs_ats(struct mlx5_ib_dev *dev,
+				       struct ib_umem *umem, int access_flags)
+{
+	if (!MLX5_CAP_GEN(dev->mdev, ats) || !umem->is_dmabuf)
+		return false;
+	return access_flags & IB_ACCESS_RELAXED_ORDERING;
+}
+
 #endif /* MLX5_IB_H */

drivers/infiniband/hw/mlx5/mr.c

@@ -935,7 +935,8 @@ static struct mlx5_ib_mr *alloc_cacheable_mr(struct ib_pd *pd,
 	 * cache then synchronously create an uncached one.
 	 */
 	if (!ent || ent->limit == 0 ||
-	    !mlx5r_umr_can_reconfig(dev, 0, access_flags)) {
+	    !mlx5r_umr_can_reconfig(dev, 0, access_flags) ||
+	    mlx5_umem_needs_ats(dev, umem, access_flags)) {
 		mutex_lock(&dev->slow_path_mutex);
 		mr = reg_create(pd, umem, iova, access_flags, page_size, false);
 		mutex_unlock(&dev->slow_path_mutex);
@@ -1016,6 +1017,8 @@ static struct mlx5_ib_mr *reg_create(struct ib_pd *pd, struct ib_umem *umem,
 	MLX5_SET(mkc, mkc, translations_octword_size,
 		 get_octo_len(iova, umem->length, mr->page_shift));
 	MLX5_SET(mkc, mkc, log_page_size, mr->page_shift);
+	if (mlx5_umem_needs_ats(dev, umem, access_flags))
+		MLX5_SET(mkc, mkc, ma_translation_mode, 1);
 	if (populate) {
 		MLX5_SET(create_mkey_in, in, translations_octword_actual_size,
 			 get_octo_len(iova, umem->length, mr->page_shift));