>On 2/23/23 10:35, Han Ding wrote: >> >> When use ovs-vsctl to delete vhostuserclient port while qemu destroy virtio, >> there is a deadlock between OVS main thread and the vhost-events thread. >> >> openvswitch is 2.14 and dpdk is 20.11 > >FWIW, 2.14 supposed to be used with 19.11.
Thanks for reply. The code flow path of openvswitch 3.0.1 and dpdk 21.11, dpdk 22.11 is same for destroying vhostuserclient device. So, I think there is a same bug in new version. >> >> Main thread: >> ofport_remove >> -> netdev_unref >> -> netdev_dpdk_vhost_destruct >> -> rte_vhost_driver_unregister (If fdentry is busy, circle again >> until the fdentry is not busy) >> -> fdset_try_del (fdentry is busy now, return -1. Goto again) >> >> vhost-nets thread: >> fdset_event_dispatch (set fdentry to busy. When destroy_device fuction >> return set the fdentry to no busy) >> -> vhost_user_read_cb >> -> vhost_user_msg_handler >> -> vhost_user_get_vring_base >> ->destroy_device >> -> ovsrcu_synchronize (Wait for other threads to quiesce) >> The vhost-nets thread wait for main thread to quiesce, but the main thread >> now is waiting for fdentry to no busy >> and circle all the time. >> >> Whether the ovsrcu_synchronize is necessary in destrory_device > >Yes, we have to wait for all threads to stop using this device. > >> or the rte_vhost_driver_unregister is correct in fdset_try_del ? > >CC: David and Maxime. > >Best regards, Ilya Maximets _______________________________________________ dev mailing list d...@openvswitch.org https://mail.openvswitch.org/mailman/listinfo/ovs-dev
