Re: [dpdk-dev] [PATCH v15] app/testpmd: support multi-process

[Andrew Rybchenko <andrew.rybchenko@oktetlabs.ru>]

On 7/8/21 3:20 PM, Min Hu (Connor) wrote:
> Hi, Andrew ,
> 
> 在 2021/7/2 20:47, Andrew Rybchenko 写道:
>> On 7/2/21 3:09 PM, Andrew Rybchenko wrote:
>>> From: "Min Hu (Connor)" <humin29@huawei.com>
>>>
>>> For example the following commands run two testpmd processes:
>>>
>>>   * the primary process:
>>>
>>> ./dpdk-testpmd --proc-type=auto -l 0-1 -- -i \
>>>     --rxq=4 --txq=4 --num-procs=2 --proc-id=0
>>>
>>>   * the secondary process:
>>>
>>> ./dpdk-testpmd --proc-type=auto -l 2-3 -- -i \
>>>     --rxq=4 --txq=4 --num-procs=2 --proc-id=1
>>>
>>> Signed-off-by: Min Hu (Connor) <humin29@huawei.com>
>>> Signed-off-by: Lijun Ou <oulijun@huawei.com>
>>> Signed-off-by: Andrew Rybchenko <Andrew.Rybchenko@oktetlabs.ru>
>>> Acked-by: Xiaoyun Li <xiaoyun.li@intel.com>
>>> Acked-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
>>> Reviewed-by: Ferruh Yigit <ferruh.yigit@intel.com>
>>
>> [snip]
>>
>>> diff --git a/app/test-pmd/testpmd.c b/app/test-pmd/testpmd.c
>>> index 1cdd3cdd1..a5da0c272 100644
>>> --- a/app/test-pmd/testpmd.c
>>> +++ b/app/test-pmd/testpmd.c
>>> @@ -520,6 +520,62 @@ enum rte_eth_rx_mq_mode rx_mq_mode =
>>> ETH_MQ_RX_VMDQ_DCB_RSS;
>>>    */
>>>   uint32_t eth_link_speed;
>>>   +/*
>>> + * ID of the current process in multi-process, used to
>>> + * configure the queues to be polled.
>>> + */
>>> +int proc_id;
>>> +
>>> +/*
>>> + * Number of processes in multi-process, used to
>>> + * configure the queues to be polled.
>>> + */
>>> +unsigned int num_procs = 1;
>>> +
>>> +static int
>>> +eth_dev_configure_mp(uint16_t port_id, uint16_t nb_rx_q, uint16_t
>>> nb_tx_q,
>>> +              const struct rte_eth_conf *dev_conf)
>>> +{
>>> +    if (is_proc_primary())
>>> +        return rte_eth_dev_configure(port_id, nb_rx_q, nb_tx_q,
>>> +                    dev_conf);
>>> +    return 0;
>>> +}
>>> +
>>> +static int
>>> +eth_dev_start_mp(uint16_t port_id)
>>> +{
>>> +    if (is_proc_primary())
>>> +        return rte_eth_dev_start(port_id);
>>> +
>>> +    return 0;
>>> +}
>>> +
>>> +static int
>>> +eth_dev_stop_mp(uint16_t port_id)
>>> +{
>>> +    if (is_proc_primary())
>>> +        return rte_eth_dev_stop(port_id);
>>> +
>>> +    return 0;
>>> +}
>>> +
>>> +static void
>>> +mempool_free_mp(struct rte_mempool *mp)
>>> +{
>>> +    if (is_proc_primary())
>>> +        rte_mempool_free(mp);
>>> +}
>>> +
>>> +static int
>>> +eth_dev_set_mtu_mp(uint16_t port_id, uint16_t mtu)
>>> +{
>>> +    if (is_proc_primary())
>>> +        return rte_eth_dev_set_mtu(port_id, mtu);
>>> +
>>> +    return 0;
>>> +}
>>> +
>>
>> I think above functions should be removed and corresponding
>> checks should be done in caller directly since above functions
>> are used in single place only and just hide what actually
>> happens in the case of secondary process. It is very
>> misleading.
>>
> This was done as Ferruh suggested in V9， and this could reduce
> the complexity for testpmd when added by the multi-process support.
>> [snip]
>>
>>> @@ -2495,21 +2565,24 @@ start_port(portid_t pid)
>>>                   return -1;
>>>               }
>>>               /* configure port */
>>> -            diag = rte_eth_dev_configure(pi, nb_rxq + nb_hairpinq,
>>> -                             nb_txq + nb_hairpinq,
>>> -                             &(port->dev_conf));
>>> +            diag = eth_dev_configure_mp(pi,
>>> +                         nb_rxq + nb_hairpinq,
>>> +                         nb_txq + nb_hairpinq,
>>> +                         &(port->dev_conf));
>>>               if (diag != 0) {
>>> -                if (rte_atomic16_cmpset(&(port->port_status),
>>> -                RTE_PORT_HANDLING, RTE_PORT_STOPPED) == 0)
>>> -                    printf("Port %d can not be set back "
>>> -                            "to stopped\n", pi);
>>> +                if (rte_atomic16_cmpset(
>>> +                        &(port->port_status),
>>> +                        RTE_PORT_HANDLING,
>>> +                        RTE_PORT_STOPPED) == 0)
>>> +                    printf("Port %d cannot be set back to stopped\n",
>>> +                        pi);
>>
>> Unrelated changes in the patch should be avoided since
>> it just makes the review harder.This will be fixed in v16.
>>
>>>                   printf("Fail to configure port %d\n", pi);
>>>                   /* try to reconfigure port next time */
>>>                   port->need_reconfig = 1;
>>>                   return -1;
>>>               }
>>>           }
>>> -        if (port->need_reconfig_queues > 0) {
>>> +        if (port->need_reconfig_queues > 0 && is_proc_primary()) {
>>>               port->need_reconfig_queues = 0;
>>>               /* setup tx queues */
>>>               for (qi = 0; qi < nb_txq; qi++) {
>>> @@ -2532,8 +2605,8 @@ start_port(portid_t pid)
>>>                   if (rte_atomic16_cmpset(&(port->port_status),
>>>                               RTE_PORT_HANDLING,
>>>                               RTE_PORT_STOPPED) == 0)
>>> -                    printf("Port %d can not be set back "
>>> -                            "to stopped\n", pi);
>>> +                    printf("Port %d cannot be set back to stopped\n",
>>> +                        pi);
>>
>> Unrelated changes in the patch should be avoided.
> This will be fixed in v16.
>>
>>>                   printf("Fail to configure port %d tx queues\n",
>>>                          pi);
>>>                   /* try to reconfigure queues next time */
>>> @@ -2610,16 +2683,16 @@ start_port(portid_t pid)
>>>           cnt_pi++;
>>>             /* start port */
>>> -        diag = rte_eth_dev_start(pi);
>>> +        diag = eth_dev_start_mp(pi);
>>>           if (diag < 0) {
>>>               printf("Fail to start port %d: %s\n", pi,
>>>                      rte_strerror(-diag));
>>>                 /* Fail to setup rx queue, return */
>>>               if (rte_atomic16_cmpset(&(port->port_status),
>>> -                RTE_PORT_HANDLING, RTE_PORT_STOPPED) == 0)
>>> -                printf("Port %d can not be set back to "
>>> -                            "stopped\n", pi);
>>> +            RTE_PORT_HANDLING, RTE_PORT_STOPPED) == 0)
>>> +                printf("Port %d cannot be set back to stopped\n",
>>> +                       pi);
>>
>> Unrelated changes in the patch should be avoided.
> This will be fixed in v16.
>>
>> [snip]
>>
>>> diff --git a/doc/guides/testpmd_app_ug/run_app.rst
>>> b/doc/guides/testpmd_app_ug/run_app.rst
>>> index eb4831835..348e5fcac 100644
>>> --- a/doc/guides/testpmd_app_ug/run_app.rst
>>> +++ b/doc/guides/testpmd_app_ug/run_app.rst
>>> @@ -545,3 +545,85 @@ The command line options are:
>>>       bit 0 - two hairpin ports loop
>>>         The default value is 0. Hairpin will use single port mode and
>>> implicit Tx flow mode.
>>> +
>>> +
>>> +Testpmd Multi-Process Command-line Options
>>> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>>> +
>>> +The following are the command-line options for testpmd multi-process
>>> support:
>>> +
>>> +*   primary process:
>>> +
>>> +.. code-block:: console
>>> +
>>> +       sudo ./dpdk-testpmd --proc-type=auto -l 0-1 -- -i --rxq=4
>>> --txq=4 \
>>> +            --num-procs=2 --proc-id=0
>>> +
>>> +*   secondary process:
>>> +
>>> +.. code-block:: console
>>> +
>>> +       sudo ./dpdk-testpmd --proc-type=auto -l 2-3 -- -i --rxq=4
>>> --txq=4 \
>>> +            --num-procs=2 --proc-id=1
>>> +
>>> +The command line options are:
>>> +
>>> +*   ``--num-procs=N``
>>> +
>>> +    The number of processes which will be used.
>>> +
>>> +*   ``--proc-id=ID``
>>> +
>>> +    The ID of the current process (ID < num-procs). ID should be
>>> different in
>>> +    primary process and secondary process, which starts from '0'.
>>> +
>>> +Calculation rule for queue:
>>> +All queues are allocated to different processes based on
>>> ``proc_num`` and
>>> +``proc_id``.
>>> +Calculation rule for the testpmd to allocate queues to each process:
>>> +
>>> +* start(queue start id) = proc_id * nb_q / num_procs
>>> +
>>> +* end(queue end id) = start + nb_q / num_procs
>>> +
>>> +For example, if testpmd is configured to have 4 Tx and Rx queues,
>>> +queues 0 and 1 will be used by the primary process and
>>> +queues 2 and 3 will be used by the secondary process.
>>> +
>>> +The number of queues should be a multiple of the number of
>>> processes. If not,
>>> +redundant queues will exist after queues are allocated to processes.
>>> If RSS
>>> +is enabled, packet loss occurs when traffic is sent to all processes
>>> at the same
>>> +time. Some traffic goes to redundant queues and cannot be forwarded.
>>> +
>>> +All the dev ops is supported in primary process. While secondary
>>> process is
>>> +not permitted to allocate or release shared memory, so some ops are
>>> not supported
>>> +as follows:
>>> +
>>> +- ``dev_configure``
>>> +- ``dev_start``
>>> +- ``dev_stop``
>>> +- ``rx_queue_setup``
>>> +- ``tx_queue_setup``
>>> +- ``rx_queue_release``
>>> +- ``tx_queue_release``
>>> +
>>> +So, any command from testpmd which calls those APIs will not be
>>> supported in
>>> +secondary process, like:
>>> +
>>> +.. code-block:: console
>>> +
>>> +    port config all rxq|txq|rxd|txd <value>
>>> +    port config <port_id> rx_offload xxx on/off
>>> +    port config <port_id> tx_offload xxx on/off
>>> +
>>> +etc.
>>
>> I did the formatting cleanup, but I still think that testpmd
>> guide should not dive into such level of details. It should
>> rather highlight multi-process behaviour specifics.
>>
>> Shouldn't testpmd store state in shared memory to avoid
>> problems when primary is stopped while secondary is running
> This could be taken into consideration in future.
> 
>>
>> Some testpmd features rely on reconfigure (i.e. simply change
>> configuration and set flag that reconfigure is required), but
>> configure does nothing and will simply ignore new settings.
>> So, it could look very-very confusing from user point of view.
>>
>> I'm not sure that it is acceptable to apply the patch in such
>> state and open huge number of bugs in testpmd behaviour when
>> multi-process is used.
>>
>> I'd even consider to exclude unsupported commands from help
>> etc. However, such level of care about user could be excessive
>> for test tool.
> This has been done in doc.
>>
>> IMHO, it should be no requirement to repeat the primary
>> process command-line configuration in the second process
>> command line (see --rxq=4 --txq=4 above). The information
>> should be obtained from shared state. In theory primary
>> process could even change some settings in interactive
> We think keeping the command line in consistent between primary
> and secondary is easy to understand for users. While shared memory for
> keeping in order or communicating could be performed,but this could
> be done in future patch.
> 
>> mode. I think testpmd should guarantee consistent behaviour
>> even in such conditions. I.e. do not allow to stop ports
>> used by forwarding running in secondary processes.
>> Run-time queues setup and deferred start should be very
>> carefully handled as well.
>  ``dev_stop`` is not allowed in secondary, which has described in doc.

I'm talking about dev_stop in primary while secondary is
running.

>>> +
>>> +Stats is supported, stats will not change when one quits and starts,
>>> as they
>>> +share the same buffer to store the stats. Flow rules are maintained
>>> in process
>>> +level: primary and secondary has its own flow list (but one flow
>>> list in HW).
>>> +The two can see all the queues, so setting the flow rules for the
>>> other is OK.
>>> +But in the testpmd primary process receiving or transmitting packets
>>> from the
>>> +queue allocated for secondary process is not permitted, and same for
>>> secondary
>>> +process.
>>> +
>>> +Flow API and RSS are supported.
>>> Thanks for your comment,
> This patch supports basic function for multi-process support in testpmd.
> I think other patches in future could enhance or optimize it, thanks.

IMHO, as I state above, current state is insufficient to
consider is a start point to be applied.