Merge tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma

Pull rdma fixes from Jason Gunthorpe:
 "Nothing particularly exciting, some small ODP regressions from the mmu
  notifier rework, another bunch of syzkaller fixes, and a bug fix for a
  botched syzkaller fix in the first rc pull request.

   - Fix busted syzkaller fix in 'get_new_pps' - this turned out to
     crash on certain HW configurations

   - Bug fixes for various missed things in error unwinds

   - Add a missing rcu_read_lock annotation in hfi/qib

   - Fix two ODP related regressions from the recent mmu notifier
     changes

   - Several more syzkaller bugs in siw, RDMA netlink, verbs and iwcm

   - Revert an old patch in CMA as it is now shown to not be allocating
     port numbers properly"

* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma:
  RDMA/iwcm: Fix iwcm work deallocation
  RDMA/siw: Fix failure handling during device creation
  RDMA/nldev: Fix crash when set a QP to a new counter but QPN is missing
  RDMA/odp: Ensure the mm is still alive before creating an implicit child
  RDMA/core: Fix protection fault in ib_mr_pool_destroy
  IB/mlx5: Fix implicit ODP race
  IB/hfi1, qib: Ensure RCU is locked when accessing list
  RDMA/core: Fix pkey and port assignment in get_new_pps
  RMDA/cm: Fix missing ib_cm_destroy_id() in ib_cm_insert_listen()
  RDMA/rw: Fix error flow during RDMA context initialization
  RDMA/core: Fix use of logical OR in get_new_pps
  Revert "RDMA/cma: Simplify rdma_resolve_addr() error flow"

2 files changed, 8 insertions, 10 deletions

diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h
index d9bffcc93587..bb78142bca5e 100644
--- a/drivers/infiniband/hw/mlx5/mlx5_ib.h
+++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h
@@ -636,6 +636,7 @@ struct mlx5_ib_mr {
 
 	/* For ODP and implicit */
 	atomic_t		num_deferred_work;
+	wait_queue_head_t       q_deferred_work;
 	struct xarray		implicit_children;
 	union {
 		struct rcu_head rcu;
diff --git a/drivers/infiniband/hw/mlx5/odp.c b/drivers/infiniband/hw/mlx5/odp.c
index 4216814ba871..bf50cd91f472 100644
--- a/drivers/infiniband/hw/mlx5/odp.c
+++ b/drivers/infiniband/hw/mlx5/odp.c
@@ -235,7 +235,8 @@ static void free_implicit_child_mr(struct mlx5_ib_mr *mr, bool need_imr_xlt)
 	mr->parent = NULL;
 	mlx5_mr_cache_free(mr->dev, mr);
 	ib_umem_odp_release(odp);
-	atomic_dec(&imr->num_deferred_work);
+	if (atomic_dec_and_test(&imr->num_deferred_work))
+		wake_up(&imr->q_deferred_work);
 }
 
 static void free_implicit_child_mr_work(struct work_struct *work)
@@ -554,6 +555,7 @@ struct mlx5_ib_mr *mlx5_ib_alloc_implicit_mr(struct mlx5_ib_pd *pd,
 	imr->umem = &umem_odp->umem;
 	imr->is_odp_implicit = true;
 	atomic_set(&imr->num_deferred_work, 0);
+	init_waitqueue_head(&imr->q_deferred_work);
 	xa_init(&imr->implicit_children);
 
 	err = mlx5_ib_update_xlt(imr, 0,
@@ -611,10 +613,7 @@ void mlx5_ib_free_implicit_mr(struct mlx5_ib_mr *imr)
 	 * 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.
 	 */
-	if (atomic_read(&imr->num_deferred_work)) {
-		flush_workqueue(system_unbound_wq);
-		WARN_ON(atomic_read(&imr->num_deferred_work));
-	}
+	wait_event(imr->q_deferred_work, !atomic_read(&imr->num_deferred_work));
 
 	/*
 	 * Fence the imr before we destroy the children. This allows us to
@@ -645,10 +644,7 @@ void mlx5_ib_fence_odp_mr(struct mlx5_ib_mr *mr)
 	/* Wait for all running page-fault handlers to finish. */
 	synchronize_srcu(&mr->dev->odp_srcu);
 
-	if (atomic_read(&mr->num_deferred_work)) {
-		flush_workqueue(system_unbound_wq);
-		WARN_ON(atomic_read(&mr->num_deferred_work));
-	}
+	wait_event(mr->q_deferred_work, !atomic_read(&mr->num_deferred_work));
 
 	dma_fence_odp_mr(mr);
 }
@@ -1720,7 +1716,8 @@ static void destroy_prefetch_work(struct prefetch_mr_work *work)
 	u32 i;
 
 	for (i = 0; i < work->num_sge; ++i)
-		atomic_dec(&work->frags[i].mr->num_deferred_work);
+		if (atomic_dec_and_test(&work->frags[i].mr->num_deferred_work))
+			wake_up(&work->frags[i].mr->q_deferred_work);
 	kvfree(work);
 }