[Bug 918] RTE EAL is unable to complete its work

[bugzilla at dpdk.org]

https://bugs.dpdk.org/show_bug.cgi?id=918

            Bug ID: 918
           Summary: RTE EAL is unable to complete its work
           Product: DPDK
           Version: 20.11
          Hardware: All
                OS: All
            Status: UNCONFIRMED
          Severity: normal
          Priority: Normal
         Component: core
          Assignee: dev at dpdk.org
          Reporter: interferation3 at gmail.com
  Target Milestone: ---

Hi guys,

I am one of two authors of Dataplane SDK, where a lot of C-based OOP modules.
Part of them is related to DPDK.

For example:

sb_class.dpdk.eal
sb_class.dpdk.lpm
sb_class.dpdk.mempool
sb_class.dpdk.ring
sb_class.netdev.dpdk
sb_class.netdev.dpdk.eth
sb_class.netdev.dpdk.eth.bonding
sb_class.netdev.dpdk.pcap
sb_class.netdev.dpdk.tap
sb_class.task.node.group.dpdk.eth
sb_class.task.node.group.dpdk.tap

I also have a module sb_class.hal (Hardware Abstraction Layer) his object runs
before sb_object.dpdk.eal and its task is to determine the hardware and
initialize the synonym table between DPDK device number and SDK object name.

Also sb_object.hal calls /usr/bin/dpdk-devbind.py script in order to connect
network cards to the required drivers.

that is, the order of initialization is as follows:

sb_object.some_project       - children initialization (they below)
sb_object.hal                - hardware detection, bind drivers
sb_object.dpdk.eal           - rte_eal_init
sb_object.dpdk.mempool       - rte_pktmbuf_pool_create
sb_object.netdev.dpdk.eth    - rte_eth_dev_configure queues etc
sb_object.lcore.scheduler.roundrobin - sb_lcore_launch
*** other objects ***

the application terminates in reverse order

and this is what I noticed when sb_object.dpdk.eal exits, I call the following
code (on the lcore that started everything)

void IMPLEMENTATION (object_free) (
    sb_status_t * a_out_err,
    sb_class_t * a_class,
    sb_object_t * a_obj)
{
    sb_object_dpdk_eal_t * obj
        = (sb_object_dpdk_eal_t *) a_obj;

    sb_object_super_class_free (a_out_err, a_class, a_obj);

    obj-> eal_cmd_line [0] = '\ 0';

#if RTE_VERSION_NUM (19, 2, 0, 0) <= RTE_VERSION
    {
        unsigned lcore_id;
        // wait for threads to stop
        // RTE_LCORE_FOREACH_WORKER (lcore_id) {
        RTE_LCORE_FOREACH (lcore_id) {
            rte_eal_wait_lcore (lcore_id);
        }
        // clean up the EAL
        rte_eal_cleanup ();
    }
#endif
}

and in gdb I still see that DPDK has not finished working and its threads are
still running.

how can I stop DPDK completely?

because when object_free of sb_object.hal is called, it attempts to bind
network cards back to the kernel stack, but DPDK is still running and the
parent would be killed as a result thole application would be marked as
"defunct" and only hardware reboot can help

Please tell me how to correctly shutdown DPDK so that:
all DPDK resources were freed
all threads created by DPDK were closed

Regards, screenshots attached


Hi guys,

I am one of two authors of Dataplane SDK, where a lot of C-based OOP modules.
Part of them is related to DPDK.

For example:

sb_class.dpdk.eal
sb_class.dpdk.lpm
sb_class.dpdk.mempool
sb_class.dpdk.ring
sb_class.netdev.dpdk
sb_class.netdev.dpdk.eth
sb_class.netdev.dpdk.eth.bonding
sb_class.netdev.dpdk.pcap
sb_class.netdev.dpdk.tap
sb_class.task.node.group.dpdk.eth
sb_class.task.node.group.dpdk.tap

I also have a module sb_class.hal (Hardware Abstraction Layer) his object runs
before sb_object.dpdk.eal and its task is to determine the hardware and
initialize the synonym table between DPDK device number and SDK object name.

Also sb_object.hal calls /usr/bin/dpdk-devbind.py script in order to connect
network cards to the required drivers.

that is, the order of initialization is as follows:

sb_object.some_project       - children initialization (they below)
sb_object.hal                - hardware detection, bind drivers
sb_object.dpdk.eal           - rte_eal_init
sb_object.dpdk.mempool       - rte_pktmbuf_pool_create
sb_object.netdev.dpdk.eth    - rte_eth_dev_configure queues etc
sb_object.lcore.scheduler.roundrobin - sb_lcore_launch
*** other objects ***

the application terminates in reverse order

and this is what I noticed when sb_object.dpdk.eal exits, I call the following
code (on the lcore that started everything)

void IMPLEMENTATION (object_free) (
    sb_status_t * a_out_err,
    sb_class_t * a_class,
    sb_object_t * a_obj)
{
    sb_object_dpdk_eal_t * obj
        = (sb_object_dpdk_eal_t *) a_obj;

    sb_object_super_class_free (a_out_err, a_class, a_obj);

    obj-> eal_cmd_line [0] = '\ 0';

#if RTE_VERSION_NUM (19, 2, 0, 0) <= RTE_VERSION
    {
        unsigned lcore_id;
        // wait for threads to stop
        // RTE_LCORE_FOREACH_WORKER (lcore_id) {
        RTE_LCORE_FOREACH (lcore_id) {
            rte_eal_wait_lcore (lcore_id);
        }
        // clean up the EAL
        rte_eal_cleanup ();
    }
#endif
}

and in gdb I still see that DPDK has not finished working and its threads are
still running.

how can I stop DPDK completely?

because when object_free of sb_object.hal is called, it attempts to bind
network cards back to the kernel stack, but DPDK is still running and the
parent would be killed as a result thole application would be marked as
"defunct" and only hardware reboot can help

Please tell me how to correctly shutdown DPDK so that:
all DPDK resources were freed
all threads created by DPDK were closed

Regards, screenshots attached

-- 
You are receiving this mail because:
You are the assignee for the bug.