RDMA/mlx5: Cleanup the synchronize_srcu() from the ODP flow

Cleanup the synchronize_srcu() from the ODP flow as it was found to be a
very heavy time consumer as part of dereg_mr.

For example de-registration of 10000 ODP MRs each with size of 2M hugepage
took 19.6 sec comparing de-registration of same number of non ODP MRs that
took 172 ms.

The new locking scheme uses the wait_event() mechanism which follows the
use count of the MR instead of using synchronize_srcu().

By that change, the time required for the above test took 95 ms which is
even better than the non ODP flow.

Once fully dropped the srcu usage, had to come with a lock to protect the
XA access.

As part of using the above mechanism we could also clean the
num_deferred_work stuff and follow the use count instead.

Link: https://lore.kernel.org/r/20210202071309.2057998-1-leon@kernel.org
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>

1 files changed, 9 insertions, 17 deletions

diff --git a/drivers/infiniband/hw/mlx5/mr.c b/drivers/infiniband/hw/mlx5/mr.c
index a63ef7c66e38..db05b0e0a8d7 100644
--- a/drivers/infiniband/hw/mlx5/mr.c
+++ b/drivers/infiniband/hw/mlx5/mr.c
@@ -158,6 +158,7 @@ static void create_mkey_callback(int status, struct mlx5_async_work *context)
 	mr->mmkey.type = MLX5_MKEY_MR;
 	mr->mmkey.key |= mlx5_idx_to_mkey(
 		MLX5_GET(create_mkey_out, mr->out, mkey_index));
+	init_waitqueue_head(&mr->mmkey.wait);
 
 	WRITE_ONCE(dev->cache.last_add, jiffies);
 
@@ -1551,10 +1552,7 @@ static struct ib_mr *create_user_odp_mr(struct ib_pd *pd, u64 start, u64 length,
 	}
 
 	odp->private = mr;
-	init_waitqueue_head(&mr->q_deferred_work);
-	atomic_set(&mr->num_deferred_work, 0);
-	err = xa_err(xa_store(&dev->odp_mkeys, mlx5_base_mkey(mr->mmkey.key),
-			      &mr->mmkey, GFP_KERNEL));
+	err = mlx5r_store_odp_mkey(dev, &mr->mmkey);
 	if (err)
 		goto err_dereg_mr;
 
@@ -1651,10 +1649,7 @@ struct ib_mr *mlx5_ib_reg_user_mr_dmabuf(struct ib_pd *pd, u64 offset,
 
 	atomic_add(ib_umem_num_pages(mr->umem), &dev->mdev->priv.reg_pages);
 	umem_dmabuf->private = mr;
-	init_waitqueue_head(&mr->q_deferred_work);
-	atomic_set(&mr->num_deferred_work, 0);
-	err = xa_err(xa_store(&dev->odp_mkeys, mlx5_base_mkey(mr->mmkey.key),
-			      &mr->mmkey, GFP_KERNEL));
+	err = mlx5r_store_odp_mkey(dev, &mr->mmkey);
 	if (err)
 		goto err_dereg_mr;
 
@@ -2330,9 +2325,7 @@ int mlx5_ib_alloc_mw(struct ib_mw *ibmw, struct ib_udata *udata)
 	}
 
 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) {
-		err = xa_err(xa_store(&dev->odp_mkeys,
-				      mlx5_base_mkey(mw->mmkey.key), &mw->mmkey,
-				      GFP_KERNEL));
+		err = mlx5r_store_odp_mkey(dev, &mw->mmkey);
 		if (err)
 			goto free_mkey;
 	}
@@ -2352,14 +2345,13 @@ int mlx5_ib_dealloc_mw(struct ib_mw *mw)
 	struct mlx5_ib_dev *dev = to_mdev(mw->device);
 	struct mlx5_ib_mw *mmw = to_mmw(mw);
 
-	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) {
-		xa_erase(&dev->odp_mkeys, mlx5_base_mkey(mmw->mmkey.key));
+	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING) &&
+	    xa_erase(&dev->odp_mkeys, mlx5_base_mkey(mmw->mmkey.key)))
 		/*
-		 * pagefault_single_data_segment() may be accessing mmw under
-		 * SRCU if the user bound an ODP MR to this MW.
+		 * pagefault_single_data_segment() may be accessing mmw
+		 * if the user bound an ODP MR to this MW.
 		 */
-		synchronize_srcu(&dev->odp_srcu);
-	}
+		mlx5r_deref_wait_odp_mkey(&mmw->mmkey);
 
 	return mlx5_core_destroy_mkey(dev->mdev, &mmw->mmkey);
 }