RDMA/mlx5: Prevent prefetch from racing with implicit destruction

Prefetch work in mlx5_ib_prefetch_mr_work can be queued and able to run
concurrently with destruction of the implicit MR. The num_deferred_work
was intended to serialize this, but there is a race:

       CPU0                                          CPU1

    mlx5_ib_free_implicit_mr()
      xa_erase(odp_mkeys)
      synchronize_srcu()
      __xa_erase(implicit_children)
                                      mlx5_ib_prefetch_mr_work()
                                        pagefault_mr()
                                         pagefault_implicit_mr()
                                          implicit_get_child_mr()
                                           xa_cmpxchg()
                                        atomic_dec_and_test(num_deferred_mr)
      wait_event(imr->q_deferred_work)
      ib_umem_odp_release(odp_imr)
        kfree(odp_imr)

At this point in mlx5_ib_free_implicit_mr() the implicit_children list is
supposed to be empty forever so that destroy_unused_implicit_child_mr()
and related are not and will not be running.

Since it is not empty the destroy_unused_implicit_child_mr() flow ends up
touching deallocated memory as mlx5_ib_free_implicit_mr() already tore down the
imr parent.

The solution is to flush out the prefetch wq by driving num_deferred_work
to zero after creation of new prefetch work is blocked.

Fixes: 5256edc ("RDMA/mlx5: Rework implicit ODP destroy")
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Leon Romanovsky <[email protected]>
Signed-off-by: Jason Gunthorpe <[email protected]>

Author: jgunthorpe

drivers/infiniband/hw/mlx5/odp.c

@@ -601,6 +601,23 @@ void mlx5_ib_free_implicit_mr(struct mlx5_ib_mr *imr)
 	 */
 	synchronize_srcu(&dev->odp_srcu);
 
+	/*
+	 * All work on the prefetch list must be completed, xa_erase() prevented
+	 * new work from being created.
+	 */
+	wait_event(imr->q_deferred_work, !atomic_read(&imr->num_deferred_work));
+
+	/*
+	 * At this point it is forbidden for any other thread to enter
+	 * pagefault_mr() on this imr. It is already forbidden to call
+	 * pagefault_mr() on an implicit child. Due to this additions to
+	 * implicit_children are prevented.
+	 */
+
+	/*
+	 * Block destroy_unused_implicit_child_mr() from incrementing
+	 * num_deferred_work.
+	 */
 	xa_lock(&imr->implicit_children);
 	xa_for_each (&imr->implicit_children, idx, mtt) {
 		__xa_erase(&imr->implicit_children, idx);
@@ -609,9 +626,8 @@ void mlx5_ib_free_implicit_mr(struct mlx5_ib_mr *imr)
 	xa_unlock(&imr->implicit_children);
 
 	/*
-	 * num_deferred_work can only be incremented inside the odp_srcu, or
-	 * under xa_lock while the child is in the xarray. Thus at this point
-	 * it is only decreasing, and all work holding it is now on the wq.
+	 * Wait for any concurrent destroy_unused_implicit_child_mr() to
+	 * complete.
 	 */
 	wait_event(imr->q_deferred_work, !atomic_read(&imr->num_deferred_work));