[PATCH v4] app/testpmd: fix primary process not polling all queues

[Jie Hai]

On 2023/6/23 0:40, Ali Alnubani wrote:
>> -----Original Message-----
>> From: Jie Hai <haijie1 at huawei.com>
>> Sent: Friday, June 9, 2023 12:04 PM
>> To: Aman Singh <aman.deep.singh at intel.com>; Yuying Zhang
>> <yuying.zhang at intel.com>; Anatoly Burakov <anatoly.burakov at intel.com>;
>> Matan Azrad <matan at nvidia.com>; Dmitry Kozlyuk
>> <dmitry.kozliuk at gmail.com>
>> Cc: dev at dpdk.org; liudongdong3 at huawei.com; shiyangx.he at intel.com;
>> ferruh.yigit at amd.com
>> Subject: [PATCH v4] app/testpmd: fix primary process not polling all queues
>>
>> Here's how the problem arises.
>> step1: Start the app.
>>      dpdk-testpmd -a 0000:35:00.0 -l 0-3 -- -i --rxq=10 --txq=10
>>
>> step2: Perform the following steps and send traffic. As expected,
>> queue 7 does not send or receive packets, and other queues do.
>>      port 0 rxq 7 stop
>>      port 0 txq 7 stop
>>      set fwd mac
>>      start
>>
>> step3: Perform the following steps and send traffic. All queues
>> are expected to send and receive packets normally, but that's not
>> the case for queue 7.
>>      stop
>>      port stop all
>>      port start all
>>      start
>>      show port xstats all
>>
>> In fact, only the value of rx_q7_packets for queue 7 is not zero,
>> which means queue 7 is enabled for the driver but is not involved
>> in packet receiving and forwarding by software. If we check queue
>> state by command 'show rxq info 0 7' and 'show txq info 0 7',
>> we see queue 7 is started as other queues are.
>>      Rx queue state: started
>>      Tx queue state: started
>> The queue 7 is started but cannot forward. That's the problem.
>>
>> We know that each stream has a read-only "disabled" field that
>> control if this stream should be used to forward. This field
>> depends on testpmd local queue state, please see
>> commit 3c4426db54fc ("app/testpmd: do not poll stopped queues").
>> DPDK framework maintains ethdev queue state that drivers reported,
>> which indicates the real state of queues.
>>
>> There are commands that update these two kind queue state such as
>> 'port X rxq|txq start|stop'. But these operations take effect only
>> in one stop-start round. In the following stop-start round, the
>> preceding operations do not take effect anymore. However, only
>> the ethdev queue state is updated, causing the testpmd and ethdev
>> state information to diverge and causing unexpected side effects
>> as above problem.
>>
>> There was a similar problem for the secondary process, please see
>> commit 5028f207a4fa ("app/testpmd: fix secondary process packet
>> forwarding").
>>
>> This patch applies its workaround with some difference to the
>> primary process. Not all PMDs implement rte_eth_rx_queue_info_get and
>> rte_eth_tx_queue_info_get, however they may support deferred_start
>> with primary process. To not break their behavior, retain the original
>> testpmd local queue state for those PMDs.
>>
>> Fixes: 3c4426db54fc ("app/testpmd: do not poll stopped queues")
>> Cc: stable at dpdk.org
>>
>> Signed-off-by: Jie Hai <haijie1 at huawei.com>
>> ---
> 
> Hi Jie,
> 
> I see the error below when starting a representor port after reattaching it with this patch, is it expected?
> 
> $ sudo ./build /app/dpdk-testpmd -n 4  -a 0000:08:00.0,dv_esw_en=1,representor=vf0-1  -a auxiliary: -a 00:00.0 --iova-mode="va" -- -i
> [..]
> testpmd> port stop all
> testpmd> port close 0
> testpmd> device detach 0000:08:00.0
> testpmd> port attach 0000:08:00.0,dv_esw_en=1,representor=0-1
> testpmd> port start 1
> Configuring Port 1 (socket 0)
> Port 1: FA:9E:D8:5F:D7:D8
> Invalid Rx queue_id=0
> testpmd: Failed to get rx queue info
> Invalid Tx queue_id=0
> testpmd: Failed to get tx queue info
> 
> Regards,
> Ali
Hi Ali,
Thanks for your feedback.

When update_queue_state is called, the status of all queues on all ports 
are updated.
The number of queues is nb_rxq|nb_txq which is stored locally by testpmd 
process.
All ports on the same process shares the same nb_rxq|nb_txq.

After detached and attached, the number of queues of port 0 is 0.
And it changes only when the port is reconfigured by testpmd,
which is when port 0 is started.

If we start port 1 first, update_queue_state will update nb_rxq|nb_txq
queues state of port 0, and that's invalid because there's zero queues.

If this patch is not applied, the same problem occurs when the secondary 
process detaches and attaches the port, and then starts the port in the 
multi-process scenario.

I will submit a patch to fix this problem. When port starts, update 
queue state based on the number of queues reported by the driver.

Thanks,
Jie Hai