RDMA/mlx5: Reduce locking in implicit_mr_get_data()

Now that the child MRs are stored in an xarray we can rely on the SRCU
lock to protect the xa_load and use xa_cmpxchg on the slow allocation path
to resolve races with concurrent page fault.

This reduces the scope of the critical section of umem_mutex for implicit
MRs to only cover mlx5_ib_update_xlt, and avoids taking a lock at all if
the child MR is already in the xarray. This makes it consistent with the
normal ODP MR critical section for umem_lock, and the locking approach
used for destroying an unusued implicit child MR.

The MLX5_IB_UPD_XLT_ATOMIC is no longer needed in implicit_get_child_mr()
since it is no longer called with any locks.

Link: https://lore.kernel.org/r/20191009160934.3143-11-jgg@ziepe.ca
Reviewed-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>

drivers/infiniband/hw/mlx5/odp.c

@@ -381,8 +381,7 @@ static struct mlx5_ib_mr *implicit_get_child_mr(struct mlx5_ib_mr *imr,
 				 MLX5_IMR_MTT_ENTRIES,
 				 PAGE_SHIFT,
 				 MLX5_IB_UPD_XLT_ZAP |
-				 MLX5_IB_UPD_XLT_ENABLE |
-				 MLX5_IB_UPD_XLT_ATOMIC);
+				 MLX5_IB_UPD_XLT_ENABLE);
 	if (err) {
 		ret = ERR_PTR(err);
 		goto out_release;
@@ -392,9 +391,16 @@ static struct mlx5_ib_mr *implicit_get_child_mr(struct mlx5_ib_mr *imr,
 	 * Once the store to either xarray completes any error unwind has to
 	 * use synchronize_srcu(). Avoid this with xa_reserve()
 	 */
-	err = xa_err(xa_store(&imr->implicit_children, idx, mr, GFP_KERNEL));
-	if (err) {
-		ret = ERR_PTR(err);
+	ret = xa_cmpxchg(&imr->implicit_children, idx, NULL, mr, GFP_KERNEL);
+	if (unlikely(ret)) {
+		if (xa_is_err(ret)) {
+			ret = ERR_PTR(xa_err(ret));
+			goto out_release;
+		}
+		/*
+		 * Another thread beat us to creating the child mr, use
+		 * theirs.
+		 */
 		goto out_release;
 	}
 
@@ -424,7 +430,8 @@ static struct mlx5_ib_mr *implicit_mr_get_data(struct mlx5_ib_mr *imr,
 	struct mlx5_ib_mr *result = NULL;
 	int ret;
 
-	mutex_lock(&odp_imr->umem_mutex);
+	lockdep_assert_held(&imr->dev->odp_srcu);
+
 	for (idx = idx; idx <= end_idx; idx++) {
 		struct mlx5_ib_mr *mtt = xa_load(&imr->implicit_children, idx);
 
@@ -450,20 +457,27 @@ static struct mlx5_ib_mr *implicit_mr_get_data(struct mlx5_ib_mr *imr,
 	 */
 out:
 	if (likely(!inv_len))
-		goto out_unlock;
+		return result;
 
+	/*
+	 * Notice this is not strictly ordered right, the KSM is updated after
+	 * the implicit_leaves is updated, so a parallel page fault could see
+	 * a MR that is not yet visible in the KSM.  This is similar to a
+	 * parallel page fault seeing a MR that is being concurrently removed
+	 * from the KSM. Both of these improbable situations are resolved
+	 * safely by resuming the HW and then taking another page fault. The
+	 * next pagefault handler will see the new information.
+	 */
+	mutex_lock(&odp_imr->umem_mutex);
 	ret = mlx5_ib_update_xlt(imr, inv_start_idx, inv_len, 0,
 				 MLX5_IB_UPD_XLT_INDIRECT |
 					 MLX5_IB_UPD_XLT_ATOMIC);
+	mutex_unlock(&odp_imr->umem_mutex);
 	if (ret) {
 		mlx5_ib_err(to_mdev(imr->ibmr.pd->device),
 			    "Failed to update PAS\n");
-		result = ERR_PTR(ret);
-		goto out_unlock;
+		return ERR_PTR(ret);
 	}
-
-out_unlock:
-	mutex_unlock(&odp_imr->umem_mutex);
 	return result;
 }