[dpdk-dev] [PATCH] examples/bond: wait for slaves to become active

[Chas Williams]

On 11/28/18 11:04 AM, Radu Nicolau wrote:
> 
> 
> On 11/28/2018 2:28 PM, Chas Williams wrote:
>>
>>
>> On 11/28/2018 08:48 AM, Radu Nicolau wrote:
>>> Hi
>>>
>>>
>>> On 11/28/2018 11:08 AM, Ferruh Yigit wrote:
>>>> On 11/14/2018 12:19 PM, Radu Nicolau wrote:
>>>>> Do not start the packet processing threads until all configured
>>>>> slaves become active.
>>>> Hi Radu,
>>>>
>>>> What happens if packet processing threads started before all slaves 
>>>> active? Exit
>>>> app, error, crash?
>>>>
>>>> So can we say this patch is fixing packet forwarding? (fix in title?)
>>>>
>>>> And do we know what break it, why this was not required previously 
>>>> but required
>>>> now? (Fixes line ?)
>>>  From what I see, the problem was always there: bond_ethdev_rx_burst 
>>> will cycle through slaves, but if called more times with no active 
>>> slaves the active slave index will point out of bounds, resulting in 
>>> a segfault.
>>> While this may require a better fix, this patch is an improvement 
>>> even if that fix comes - the configured slaves needs to be checked, 
>>> and if none became active there is no point going further.
>>>
>>> in bond_ethdev_rx_burst:
>>>
>>> slave_count = internals->active_slave_count;
>>> ...
>>>      if (++internals->active_slave == slave_count)
>>>          internals->active_slave = 0;
>>> slave_count is zero, the if() will never be true and active_slave 
>>> will be continuously incremented. It was not written to work with no 
>>> active slaves.
>>
>> Just create another patch for the rx routines.  If the active_slave_count
>> is 0, there's nothing to do really.  It should just return and not
>> bother with any of the other work.
> I can do that, and it will be the better fix I mentioned.
> But I still think this patch makes the sample app better, at least it 
> gives a hint to someone looking to develop its own app to check on the 
> slaves' status before proceeding to rx.

Yes, I agree this patch is still valid. If you are writing some sort of 
test, you should wait until the bonding interface and slaves are ready 
before you start sending traffic.

>>
>>>
>>>>
>>>> Thanks,
>>>> ferruh
>>>>
>>>>> Signed-off-by: Radu Nicolau <radu.nicolau at intel.com>
>>>>> ---
>>>>>   examples/bond/main.c | 15 +++++++++++++++
>>>>>   1 file changed, 15 insertions(+)
>>>>>
>>>>> diff --git a/examples/bond/main.c b/examples/bond/main.c
>>>>> index b282e68..6623cae 100644
>>>>> --- a/examples/bond/main.c
>>>>> +++ b/examples/bond/main.c
>>>>> @@ -220,6 +220,7 @@ bond_port_init(struct rte_mempool *mbuf_pool)
>>>>>       struct rte_eth_rxconf rxq_conf;
>>>>>       struct rte_eth_txconf txq_conf;
>>>>>       struct rte_eth_conf local_port_conf = port_conf;
>>>>> +    uint16_t wait_counter = 20;
>>>>>       retval = rte_eth_bond_create("net_bonding0", BONDING_MODE_ALB,
>>>>>               0 /*SOCKET_ID_ANY*/);
>>>>> @@ -274,6 +275,20 @@ bond_port_init(struct rte_mempool *mbuf_pool)
>>>>>       if (retval < 0)
>>>>>           rte_exit(retval, "Start port %d failed (res=%d)", 
>>>>> BOND_PORT, retval);
>>>>> +    printf("Waiting for slaves to become active...");
>>>>> +    while (wait_counter) {
>>>>> +        uint16_t act_slaves[16] = {0};
>>>>> +        if (rte_eth_bond_active_slaves_get(BOND_PORT, act_slaves, 
>>>>> 16) ==
>>>>> +                slaves_count) {
>>>>> +            printf("\n");
>>>>> +            break;
>>>>> +        }
>>>>> +        sleep(1);
>>>>> +        printf("...");
>>>>> +        if (--wait_counter == 0)
>>>>> +            rte_exit(-1, "\nFailed to activate slaves\n");
>>>>> +    }
>>>>> +
>>>>>       rte_eth_promiscuous_enable(BOND_PORT);
>>>>>       struct ether_addr addr;
>>>>>
>>>
>