Hi,team.

During the forwarding process, we used bpf to capture
the pmd cutout time and found that the pmd process was
locked in the upcall process and was unable to process
the forwarding process.

Here I think it is not appropriate for the pmd process
and the upcall process of ovs-dpdk to be coupled together.
There are a large number of locks in the upcall process,
which will cause the pmd to be locked and unable to send
and receive packets. Why don't we consider splitting the
upcall in the ovs-dpdk scenario to reduce the impact on pmd.


12:54:07 575095643125024 - Switch: [062] swapper/62:0 -> pmd-c62/id:135:454907 
| latency 57018 us | curr_task(swapper/62-0)
        __GI___lll_lock_wait+80
        __GI___pthread_mutex_lock+268
        ovs_mutex_lock_at+32
        seq_change+40
        tnl_neigh_set+208
        compose_output_action__+2712
        output_normal+584
        xlate_output_action+2524
        do_xlate_actions+712
        xlate_actions+2764
        process_upcall+852
        upcall_cb+304
        dp_netdev_upcall+172
        fast_path_processing+2580
        dp_netdev_input__+2652
        dp_netdev_input+20
        dp_netdev_process_rxq_port+352
        0xffffffff0000
        ovsthread_wrapper+136
        start_thread+724
        thread_start+12

21:42:05 520373736472757 - Switch: [095] swapper/95:0 -> pmd-c95/id:133:454905 
| latency 13706 us | curr_task(swapper/95-0)
        __GI___lll_lock_wait+80
        __GI___pthread_mutex_lock+268
        ovs_mutex_lock_at+32
        fat_rwlock_rdlock+56
        tnl_port_send+52
        compose_output_action__+1700
        xlate_output_action+1252
        do_xlate_actions+8620
        do_xlate_actions+5264
        ......
        do_xlate_actions+5264
        clone_xlate_actions+316
        do_xlate_actions+4156
        do_xlate_actions+5264
        ......
        do_xlate_actions+5264
        xlate_actions+2764
        process_upcall+852
        upcall_cb+304
        dp_netdev_upcall+172
        fast_path_processing+2580
        dp_netdev_input__+2652
        dp_netdev_input+20
        dp_netdev_process_rxq_port+352
        0xffffffff0000
        ovsthread_wrapper+136
        start_thread+724
        thread_start+12

21:42:05 520373736243062 - Switch: [094] swapper/94:0 -> pmd-c94/id:137:454909 
| latency 13992 us | curr_task(swapper/94-0)
        __GI___lll_lock_wait+80
        __GI___pthread_mutex_lock+268
        ovs_mutex_lock_at+32
        fat_rwlock_rdlock+56
        tnl_port_receive+76
        xlate_lookup+124
        upcall_cb+188
        dp_netdev_upcall+172
        fast_path_processing+2580
        dp_netdev_input__+2652
        dp_execute_cb+4040
        odp_execute_actions+452
        dp_netdev_input__+1204
        dp_netdev_input+20
        dp_netdev_process_rxq_port+352
        0xffffffff0000
        ovsthread_wrapper+136
        start_thread+724
        thread_start+12

15:56:12 499620650630327 - Switch: [094] swapper/94:0 -> pmd-c94/id:137:2566810 
| latency 640 us | curr_task(swapper/94-0):
        __GI___lll_lock_wait+80
        __GI___pthread_mutex_lock+712
        ovs_mutex_lock_at+32
        rule_dpif_credit_stats+48
        do_xlate_actions+5196
        do_xlate_actions+5264
         ......
        do_xlate_actions+5264
        xlate_actions+2764
        process_upcall+852
        upcall_cb+304
        dp_netdev_upcall+172
        fast_path_processing+2580
        dp_netdev_input__+2652
        dp_netdev_input+20
        dp_netdev_process_rxq_port+352
        0xffffffff0000
        ovsthread_wrapper+136
        start_thread+724
        thread_start+12



Jun Wang
_______________________________________________
discuss mailing list
disc...@openvswitch.org
https://mail.openvswitch.org/mailman/listinfo/ovs-discuss

Reply via email to