vdpa/mlx5: Postpone MR deletion

Currently, when a new MR is set up, the old MR is deleted. MR deletion
is about 30-40% the time of MR creation. As deleting the old MR is not
important for the process of setting up the new MR, this operation
can be postponed.

This series adds a workqueue that does MR garbage collection at a later
point. If the MR lock is taken, the handler will back off and
reschedule. The exception during shutdown: then the handler must
not postpone the work.

Note that this is only a speculative optimization: if there is some
mapping operation that is triggered while the garbage collector handler
has the lock taken, this operation it will have to wait for the handler
to finish.

Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Reviewed-by: Cosmin Ratiu <cratiu@nvidia.com>
Message-Id: <20240830105838.2666587-9-dtatulea@nvidia.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>

3 files changed, 64 insertions, 5 deletions

diff --git a/drivers/vdpa/mlx5/core/mlx5_vdpa.h b/drivers/vdpa/mlx5/core/mlx5_vdpa.h
index c3e17bc888e8..2cedf7e2dbc4 100644
--- a/drivers/vdpa/mlx5/core/mlx5_vdpa.h
+++ b/drivers/vdpa/mlx5/core/mlx5_vdpa.h
@@ -86,8 +86,18 @@ enum {
 struct mlx5_vdpa_mr_resources {
 	struct mlx5_vdpa_mr *mr[MLX5_VDPA_NUM_AS];
 	unsigned int group2asid[MLX5_VDPA_NUMVQ_GROUPS];
+
+	/* Pre-deletion mr list */
 	struct list_head mr_list_head;
+
+	/* Deferred mr list */
+	struct list_head mr_gc_list_head;
+	struct workqueue_struct *wq_gc;
+	struct delayed_work gc_dwork_ent;
+
 	struct mutex lock;
+
+	atomic_t shutdown;
 };
 
 struct mlx5_vdpa_dev {
diff --git a/drivers/vdpa/mlx5/core/mr.c b/drivers/vdpa/mlx5/core/mr.c
index 64683a39d3db..2dd21e0b399e 100644
--- a/drivers/vdpa/mlx5/core/mr.c
+++ b/drivers/vdpa/mlx5/core/mr.c
@@ -653,14 +653,50 @@ static void _mlx5_vdpa_destroy_mr(struct mlx5_vdpa_dev *mvdev, struct mlx5_vdpa_
 	kfree(mr);
 }
 
+/* There can be multiple .set_map() operations in quick succession.
+ * This large delay is a simple way to prevent the MR cleanup from blocking
+ * .set_map() MR creation in this scenario.
+ */
+#define MLX5_VDPA_MR_GC_TRIGGER_MS 2000
+
+static void mlx5_vdpa_mr_gc_handler(struct work_struct *work)
+{
+	struct mlx5_vdpa_mr_resources *mres;
+	struct mlx5_vdpa_mr *mr, *tmp;
+	struct mlx5_vdpa_dev *mvdev;
+
+	mres = container_of(work, struct mlx5_vdpa_mr_resources, gc_dwork_ent.work);
+
+	if (atomic_read(&mres->shutdown)) {
+		mutex_lock(&mres->lock);
+	} else if (!mutex_trylock(&mres->lock)) {
+		queue_delayed_work(mres->wq_gc, &mres->gc_dwork_ent,
+				   msecs_to_jiffies(MLX5_VDPA_MR_GC_TRIGGER_MS));
+		return;
+	}
+
+	mvdev = container_of(mres, struct mlx5_vdpa_dev, mres);
+
+	list_for_each_entry_safe(mr, tmp, &mres->mr_gc_list_head, mr_list) {
+		_mlx5_vdpa_destroy_mr(mvdev, mr);
+	}
+
+	mutex_unlock(&mres->lock);
+}
+
 static void _mlx5_vdpa_put_mr(struct mlx5_vdpa_dev *mvdev,
 			      struct mlx5_vdpa_mr *mr)
 {
+	struct mlx5_vdpa_mr_resources *mres = &mvdev->mres;
+
 	if (!mr)
 		return;
 
-	if (refcount_dec_and_test(&mr->refcount))
-		_mlx5_vdpa_destroy_mr(mvdev, mr);
+	if (refcount_dec_and_test(&mr->refcount)) {
+		list_move_tail(&mr->mr_list, &mres->mr_gc_list_head);
+		queue_delayed_work(mres->wq_gc, &mres->gc_dwork_ent,
+				   msecs_to_jiffies(MLX5_VDPA_MR_GC_TRIGGER_MS));
+	}
 }
 
 void mlx5_vdpa_put_mr(struct mlx5_vdpa_dev *mvdev,
@@ -851,9 +887,17 @@ int mlx5_vdpa_init_mr_resources(struct mlx5_vdpa_dev *mvdev)
 {
 	struct mlx5_vdpa_mr_resources *mres = &mvdev->mres;
 
-	INIT_LIST_HEAD(&mres->mr_list_head);
+	mres->wq_gc = create_singlethread_workqueue("mlx5_vdpa_mr_gc");
+	if (!mres->wq_gc)
+		return -ENOMEM;
+
+	INIT_DELAYED_WORK(&mres->gc_dwork_ent, mlx5_vdpa_mr_gc_handler);
+
 	mutex_init(&mres->lock);
 
+	INIT_LIST_HEAD(&mres->mr_list_head);
+	INIT_LIST_HEAD(&mres->mr_gc_list_head);
+
 	return 0;
 }
 
@@ -861,5 +905,10 @@ void mlx5_vdpa_destroy_mr_resources(struct mlx5_vdpa_dev *mvdev)
 {
 	struct mlx5_vdpa_mr_resources *mres = &mvdev->mres;
 
+	atomic_set(&mres->shutdown, 1);
+
+	flush_delayed_work(&mres->gc_dwork_ent);
+	destroy_workqueue(mres->wq_gc);
+	mres->wq_gc = NULL;
 	mutex_destroy(&mres->lock);
 }
diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c b/drivers/vdpa/mlx5/net/mlx5_vnet.c
index 4b7dcfcba444..dee019977716 100644
--- a/drivers/vdpa/mlx5/net/mlx5_vnet.c
+++ b/drivers/vdpa/mlx5/net/mlx5_vnet.c
@@ -3435,6 +3435,8 @@ static void mlx5_vdpa_free(struct vdpa_device *vdev)
 	free_fixed_resources(ndev);
 	mlx5_vdpa_clean_mrs(mvdev);
 	mlx5_vdpa_destroy_mr_resources(&ndev->mvdev);
+	mlx5_cmd_cleanup_async_ctx(&mvdev->async_ctx);
+
 	if (!is_zero_ether_addr(ndev->config.mac)) {
 		pfmdev = pci_get_drvdata(pci_physfn(mvdev->mdev->pdev));
 		mlx5_mpfs_del_mac(pfmdev, ndev->config.mac);
@@ -4042,8 +4044,6 @@ static void mlx5_vdpa_dev_del(struct vdpa_mgmt_dev *v_mdev, struct vdpa_device *
 	mvdev->wq = NULL;
 	destroy_workqueue(wq);
 	mgtdev->ndev = NULL;
-
-	mlx5_cmd_cleanup_async_ctx(&mvdev->async_ctx);
 }
 
 static int mlx5_vdpa_set_attr(struct vdpa_mgmt_dev *v_mdev, struct vdpa_device *dev,