Re: [dpdk-dev] [PATCH v2 1/2] app/testpmd: add link speed check before port start

[Huisong Li <lihuisong@huawei.com>]

在 2021/5/6 10:36, Li, Xiaoyun 写道:
>
>> -----Original Message-----
>> From: Huisong Li <lihuisong@huawei.com>
>> Sent: Tuesday, May 4, 2021 09:46
>> To: Li, Xiaoyun <xiaoyun.li@intel.com>
>> Cc: Yigit, Ferruh <ferruh.yigit@intel.com>; dev@dpdk.org
>> Subject: Re: [dpdk-dev] [PATCH v2 1/2] app/testpmd: add link speed check
>> before port start
>>
>>
>> 在 2021/4/30 12:46, Li, Xiaoyun 写道:
>>>> -----Original Message-----
>>>> From: Huisong Li <lihuisong@huawei.com>
>>>> Sent: Friday, April 30, 2021 12:04
>>>> To: Li, Xiaoyun <xiaoyun.li@intel.com>
>>>> Cc: Yigit, Ferruh <ferruh.yigit@intel.com>; dev@dpdk.org
>>>> Subject: Re: [dpdk-dev] [PATCH v2 1/2] app/testpmd: add link speed
>>>> check before port start
>>>>
>>>>
>>>> 在 2021/4/30 11:19, Li, Xiaoyun 写道:
>>>>>> -----Original Message-----
>>>>>> From: Min Hu (Connor) <humin29@huawei.com>
>>>>>> Sent: Wednesday, April 28, 2021 16:37
>>>>>> To: dev@dpdk.org
>>>>>> Cc: Yigit, Ferruh <ferruh.yigit@intel.com>; Li, Xiaoyun
>>>>>> <xiaoyun.li@intel.com>
>>>>>> Subject: [PATCH v2 1/2] app/testpmd: add link speed check before
>>>>>> port start
>>>>>>
>>>>>> From: Huisong Li <lihuisong@huawei.com>
>>>>>>
>>>>>> Currently, to check whether the configured link_speeds is valid, we
>>>>>> have to run "port start". In addition, if the configuration fails,
>>>>>> "port-
>>>>> dev_conf.link_speeds"
>>>>>> maintained in testpmd cannot be restored.
>>>>>>
>>>>>> This patch adds the link_speeds check before port start by calling
>>>>>> dev_configure, and resolves these problems.
>>>>> Not sure about this patch. I don't think you can fix the issue you mentioned.
>>>>> Probably only hns3 does speed check in dev_configure. I don't see
>>>>> this in other
>>>> drivers, not in i40e/ice/mlx.
>>>>> I guess it's because if it's not supported speed, it will just be
>>>>> UNKNOWN and
>>>> user can config again?
>>>>
>>>> I think that the validity of the configuration delivered by
>>>> dev_configure is ensured by this interface and cannot be left to the backend.
>>>>
>>>> Because it facilitates users to handle abnormal configurations in a
>>>> timely manner. It may be more appropriate for the driver to do this
>>>> check in dev_configure.
>>> I still think it's not necessary.
>> ok😂
>>
>> @Ferruh, what do you think?
>>
>>>> In addition, even if other drivers do not add this check in
>>>> dev_configure, this patch does not seem to affect the current behavior of
>> these drivers.
>>>>> BTW, even if this behavior is accepted by others, still some comments below.
>>>>>
>>>>>> Signed-off-by: Huisong Li <lihuisong@huawei.com>
>>>>>> Signed-off-by: Min Hu (Connor) <humin29@huawei.com>
>>>>>> ---
>>>>>>     app/test-pmd/cmdline.c | 42
>>>>>> ++++++++++++++++++++++++++++++++++++++++-
>>>>>> -
>>>>>>     1 file changed, 40 insertions(+), 2 deletions(-)
>>>>>>
>>>>>> diff --git a/app/test-pmd/cmdline.c b/app/test-pmd/cmdline.c index
>>>>>> 5fdcc1c..ddbc629 100644
>>>>>> --- a/app/test-pmd/cmdline.c
>>>>>> +++ b/app/test-pmd/cmdline.c
>>>>>> @@ -1549,8 +1549,12 @@ cmd_config_speed_all_parsed(void
>>>> *parsed_result,
>>>>>>     			__rte_unused void *data)
>>>>>>     {
>>>>>>     	struct cmd_config_speed_all *res = parsed_result;
>>>>>> +	uint32_t old_link_speeds[RTE_MAX_ETHPORTS];
>>>>>> +	struct rte_port *port;
>>>>>>     	uint32_t link_speed;
>>>>>>     	portid_t pid;
>>>>>> +	portid_t i;
>>>>>> +	int ret;
>>>>>>
>>>>>>     	if (!all_ports_stopped()) {
>>>>>>     		printf("Please stop all ports first\n"); @@ -1562,7 +1566,26
>>>>>> @@ cmd_config_speed_all_parsed(void *parsed_result,
>>>>>>     		return;
>>>>>>
>>>>>>     	RTE_ETH_FOREACH_DEV(pid) {
>>>>>> -		ports[pid].dev_conf.link_speeds = link_speed;
>>>>>> +		port = &ports[pid];
>>>>>> +		old_link_speeds[pid] = port->dev_conf.link_speeds;
>>>>>> +		port->dev_conf.link_speeds = link_speed;
>>>>>> +		ret = rte_eth_dev_configure(pid, nb_rxq, nb_txq,
>>>>>> +					    &port->dev_conf);
>>>>>> +		if (ret < 0) {
>>>>>> +			printf("Failed to check link speeds for port %d, ret
>>>>>> = %d.\n",
>>>>>> +				pid, ret);
>>>>>> +			goto roolback;
>>>>> Why don't you just add restoring all ports speed here and then
>>>>> "break"? No
>>>> matter one dev fails or not, all ports will do reconfig from your code logic.
>>>>> And you type rollback wrongly.
>>>> It cannot exit directly after restoring all ports speed. If the cmd
>>>> fails to execute, it is necessary to reconfigure device with the correct
>> configuration.
>>>>     Because "nb_rx/tx_queues" in dev->data are cleared to zero if
>>>> dev_configure fails to be executed in PMD driver.
>>> ?
>>> cmd_reconfig_device_queue(RTE_PORT_ALL, 1, 1); is at the end of this cmd.
>> This will re-config all ports.
>>> I don't understand why can't you add a restoring in this if and break this loop
>> and do this reconfig.
>> What do you mean? If a port fails among all ports, only the failed port is
>> restored and reconfigured. I suppose that's what you mean?
>> The cmd "port config all speed xxx duplex xxx" applies to all ports. If it fails to be
>> delivered, the user considers that the cmd does not take effect.
>> So it is necessary to restore and reconfigure all ports to the previous state.
> You still don't understand me. See below. I mean the following.
> You don't need to got to rollback. Because "cmd_reconfig_device_queue(RTE_PORT_ALL, 1, 1);" will config all ports anyway already no matter you restore speeds or not.
> Can you read the code carefully? "cmd_reconfig_device_queue(RTE_PORT_ALL, 1, 1);" is always there and it configs ALL ports.
>
> static void
> cmd_config_speed_all_parsed(void *parsed_result,
> 			__rte_unused struct cmdline *cl,
> 			__rte_unused void *data)
> {
> 	struct cmd_config_speed_all *res = parsed_result;
> 	uint32_t link_speed;
> 	portid_t pid;
>
> 	if (!all_ports_stopped()) {
> 		printf("Please stop all ports first\n");
> 		return;
> 	}
>
> 	if (parse_and_check_speed_duplex(res->value1, res->value2,
> 			&link_speed) < 0)
> 		return;
>
> 	RTE_ETH_FOREACH_DEV(pid) {
> -		ports[pid].dev_conf.link_speeds = link_speed;
> +		port = &ports[pid];
> +		old_link_speeds[pid] = port->dev_conf.link_speeds;
> +		port->dev_conf.link_speeds = link_speed;
> +		ret = rte_eth_dev_configure(pid, nb_rxq, nb_txq,
> +					    &port->dev_conf);
> +		if (ret < 0) {
> +			printf("Failed to check link speeds for port %d, ret = %d.\n",
> +				pid, ret);
> +			for (i = 0; i <= pid; i++) {
> +				port = &ports[i];
> +				port->dev_conf.link_speeds = old_link_speeds[i];
> +			}
> +			break;
> +		}
> 	}
>
> 	cmd_reconfig_device_queue(RTE_PORT_ALL, 1, 1);
> }

It looks good!  Now we're left with a question about whether we need to 
add this link speed check.

@Ferruh, what do you think?

>>>>>> +		}
>>>>>> +	}
>>>>>> +
>>>>>> +	cmd_reconfig_device_queue(RTE_PORT_ALL, 1, 1);
>>>>>> +
>>>>>> +	return;
>>>>>> +
>>>>>> +roolback:
>>>>>> +	for (i = 0; i <= pid; i++) {
>>>>>> +		port = &ports[i];
>>>>>> +		port->dev_conf.link_speeds = old_link_speeds[i];
>>>>>>     	}
>>>>>>
>>>>>>     	cmd_reconfig_device_queue(RTE_PORT_ALL, 1, 1); @@ -1621,7
>>>>>> +1644,10 @@ cmd_config_speed_specific_parsed(void *parsed_result,
>>>>>>     				__rte_unused void *data)
>>>>>>     {
>>>>>>     	struct cmd_config_speed_specific *res = parsed_result;
>>>>>> +	uint32_t old_link_speeds;
>>>>>> +	struct rte_port *port;
>>>>>>     	uint32_t link_speed;
>>>>>> +	int ret;
>>>>>>
>>>>>>     	if (!all_ports_stopped()) {
>>>>>>     		printf("Please stop all ports first\n"); @@ -1635,8 +1661,20
>>>>>> @@ cmd_config_speed_specific_parsed(void *parsed_result,
>>>>>>     			&link_speed) < 0)
>>>>>>     		return;
>>>>>>
>>>>>> -	ports[res->id].dev_conf.link_speeds = link_speed;
>>>>>> +	port = &ports[res->id];
>>>>>> +	old_link_speeds = port->dev_conf.link_speeds;
>>>>>> +	port->dev_conf.link_speeds = link_speed;
>>>>>> +	ret = rte_eth_dev_configure(res->id, nb_rxq, nb_txq,
>>>>>> +				    &port->dev_conf);
>>>>>> +	if (ret < 0) {
>>>>>> +		printf("Failed to check link speeds for port %d, ret = %d.\n",
>>>>>> +			res->id, ret);
>>>>>> +		port->dev_conf.link_speeds = old_link_speeds;
>>>>>> +	}
>>>>>>
>>>>>> +	/*
>>>>>> +	 * If the cmd fails to execute, it is necessary to reconfigure device.
>>>>>> +	 */
>>>>>>     	cmd_reconfig_device_queue(RTE_PORT_ALL, 1, 1);  }
>>>>>>
>>>>>> --
>>>>>> 2.7.4
>>>>> .