> -----Original Message-----
> From: Stojaczyk, DariuszX
> Sent: Tuesday, May 16, 2017 4:35 PM
> To: dev@dpdk.org
> Cc: Wodkowski, PawelX <pawelx.wodkow...@intel.com>; Stojaczyk, DariuszX
> <dariuszx.stojac...@intel.com>
> Subject: [PATCH] vhost: fix deadlock on rte_vhost_driver_unregister()
>
> Consider the following scenario, threads A and B:
> (A)
> * fdset_event_dispatch() start
> * pfdentry->busy = 1;
> * vhost_user_read_cb() start
> * vhost_destroy_device() start
> (B)
> * rte_vhost_driver_unregister() start
> * pthread_mutex_lock(&vsocket->conn_mutex);
> * fdset_del()
> * endless loop, waiting for pfdentry->busy == 0
> (A)
> * vhost_destroy_device() end
> * pthread_mutex_lock(&vsocket->conn_mutex);
> (mutex already locked - deadlock at this point)
>
> Thread B has locked vsocket->conn_mutex and is in while(1) loop waiting for
> given fd to change it's busy flag to 0.
> Thread A would have to finish vhost_user_read_cb() in order to set busy flag
> back to 0, but that can't happen due to the vsocket->conn_mutex lock.
>
> This patch synchronizes rte_vhost_driver_unregister() with
> vhost_user_read_cb() through vhost_user.mutex.
>
> Signed-off-by: Dariusz Stojaczyk <dariuszx.stojac...@intel.com>
> ---
> lib/librte_vhost/socket.c | 4 ++++
> 1 file changed, 4 insertions(+)
>
> diff --git a/lib/librte_vhost/socket.c b/lib/librte_vhost/socket.c index
> c7f99b0..77e58fe 100644
> --- a/lib/librte_vhost/socket.c
> +++ b/lib/librte_vhost/socket.c
> @@ -273,6 +273,8 @@ vhost_user_read_cb(int connfd, void *dat, int
> *remove)
>
> ret = vhost_user_msg_handler(conn->vid, connfd);
> if (ret < 0) {
> + pthread_mutex_lock(&vhost_user.mutex);
> +
> close(connfd);
> *remove = 1;
> vhost_destroy_device(conn->vid);
> @@ -287,6 +289,8 @@ vhost_user_read_cb(int connfd, void *dat, int
> *remove)
> create_unix_socket(vsocket);
> vhost_user_start_client(vsocket);
> }
> +
> + pthread_mutex_unlock(&vhost_user.mutex);
> }
> }
>
> --
> 2.7.4
I've found other cases where rte_vhost_driver_unregister() can still deadlock.
Please do not merge this patch now, I will try to come up with a different
solution for this issue.
