Re: [PATCH rdma-rc] RDMA/mlx5: Fix devlink deadlock on net namespace deletion

From: Jason Gunthorpe <jgg@nvidia.com>
To: Leon Romanovsky <leon@kernel.org>
Cc: Doug Ledford <dledford@redhat.com>,
	Parav Pandit <parav@nvidia.com>, Jakub Kicinski <kuba@kernel.org>,
	Jiri Pirko <jiri@mellanox.com>, <linux-rdma@vger.kernel.org>,
	Michael Guralnik <michaelgur@mellanox.com>,
	<netdev@vger.kernel.org>, Saeed Mahameed <saeedm@nvidia.com>
Subject: Re: [PATCH rdma-rc] RDMA/mlx5: Fix devlink deadlock on net namespace deletion
Date: Mon, 19 Oct 2020 10:07:51 -0300	[thread overview]
Message-ID: <20201019130751.GE6219@nvidia.com> (raw)
In-Reply-To: <20201019052736.628909-1-leon@kernel.org>

On Mon, Oct 19, 2020 at 08:27:36AM +0300, Leon Romanovsky wrote:
> From: Parav Pandit <parav@nvidia.com>
> 
> When a mlx5 core devlink instance is reloaded in different net
> namespace, its associated IB device is deleted and recreated.
> 
> Example sequence is:
> $ ip netns add foo
> $ devlink dev reload pci/0000:00:08.0 netns foo
> $ ip netns del foo
> 
> mlx5 IB device needs to attach and detach the netdevice to it
> through the netdev notifier chain during load and unload sequence.
> A below call graph of the unload flow.
> 
> cleanup_net()
>    down_read(&pernet_ops_rwsem); <- first sem acquired
>      ops_pre_exit_list()
>        pre_exit()
>          devlink_pernet_pre_exit()
>            devlink_reload()
>              mlx5_devlink_reload_down()
>                mlx5_unload_one()
>                [...]
>                  mlx5_ib_remove()
>                    mlx5_ib_unbind_slave_port()
>                      mlx5_remove_netdev_notifier()
>                        unregister_netdevice_notifier()
>                          down_write(&pernet_ops_rwsem);<- recurrsive lock
> 
> Hence, when net namespace is deleted, mlx5 reload results in deadlock.
> 
> When deadlock occurs, devlink mutex is also held. This not only deadlocks
> the mlx5 device under reload, but all the processes which attempt to access
> unrelated devlink devices are deadlocked.
> 
> Hence, fix this by mlx5 ib driver to register for per net netdev
> notifier instead of global one, which operats on the net namespace
> without holding the pernet_ops_rwsem.
> 
> Fixes: 4383cfcc65e7 ("net/mlx5: Add devlink reload")
> Signed-off-by: Parav Pandit <parav@nvidia.com>
> Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
>  drivers/infiniband/hw/mlx5/main.c                  | 6 ++++--
>  drivers/net/ethernet/mellanox/mlx5/core/lib/mlx5.h | 5 -----
>  include/linux/mlx5/driver.h                        | 5 +++++
>  3 files changed, 9 insertions(+), 7 deletions(-)
> 
> diff --git a/drivers/infiniband/hw/mlx5/main.c b/drivers/infiniband/hw/mlx5/main.c
> index 944bb7691913..b1b3e563c15e 100644
> +++ b/drivers/infiniband/hw/mlx5/main.c
> @@ -3323,7 +3323,8 @@ static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
>  	int err;
> 
>  	dev->port[port_num].roce.nb.notifier_call = mlx5_netdev_event;
> -	err = register_netdevice_notifier(&dev->port[port_num].roce.nb);
> +	err = register_netdevice_notifier_net(mlx5_core_net(dev->mdev),
> +					      &dev->port[port_num].roce.nb);

This looks racy, what lock needs to be held to keep *mlx5_core_net()
stable?

>  	if (err) {
>  		dev->port[port_num].roce.nb.notifier_call = NULL;
>  		return err;
> @@ -3335,7 +3336,8 @@ static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
>  static void mlx5_remove_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
>  {
>  	if (dev->port[port_num].roce.nb.notifier_call) {
> -		unregister_netdevice_notifier(&dev->port[port_num].roce.nb);
> +		unregister_netdevice_notifier_net(mlx5_core_net(dev->mdev),
> +						  &dev->port[port_num].roce.nb);

This seems dangerous too, what if the mlx5_core_net changed before we
get here?

What are the rules for when devlink_net() changes?

Jason