[dpdk-stable] [PATCH 2/3] app/testpmd: fix RSS hash type update

[Maxime Coquelin]

Hi Xiaoyun,

On 9/16/21 5:03 AM, Li, Xiaoyun wrote:
> Hi
> 
>> -----Original Message-----
>> From: stable <stable-bounces at dpdk.org> On Behalf Of Nélio Laranjeiro
>> Sent: Tuesday, September 14, 2021 15:20
>> To: Maxime Coquelin <maxime.coquelin at redhat.com>; Yigit, Ferruh
>> <ferruh.yigit at intel.com>
>> Cc: Andrew Rybchenko <andrew.rybchenko at oktetlabs.ru>; dev at dpdk.org; Xia,
>> Chenbo <chenbo.xia at intel.com>; amorenoz at redhat.com;
>> david.marchand at redhat.com; michaelba at nvidia.com; viacheslavo at nvidia.com;
>> stable at dpdk.org; Shahaf Shuler <shahafs at nvidia.com>
>> Subject: Re: [dpdk-stable] [PATCH 2/3] app/testpmd: fix RSS hash type update
>>
>> +Shahaf,
>>
>> Hi Maxime,
>>
>> On Mon, Sep 13, 2021 at 11:41:04AM +0200, Maxime Coquelin wrote:
>>> Hi Nélio,
>>>
>>> On 9/10/21 4:16 PM, Nélio Laranjeiro wrote:
>>>> On Fri, Sep 10, 2021 at 01:06:53PM +0300, Andrew Rybchenko wrote:
>>>>> On 9/10/21 12:57 PM, Maxime Coquelin wrote:
>>>>>>
>>>>>>
>>>>>> On 9/10/21 11:51 AM, Andrew Rybchenko wrote:
>>>>>>> On 9/10/21 12:17 PM, Maxime Coquelin wrote:
>>>>>>>> port_rss_hash_key_update() initializes rss_conf with the RSS
>>>>>>>> hash type and key provided by the user, but it calls
>>>>>>>> rte_eth_dev_rss_hash_conf_get() before calling
>>>>>>>> rte_eth_dev_rss_hash_update(), which overides the parsed config
>>>>>>>> with current NIC's config.
>>>>>>>>
>>>>>>>> While the RSS key value is set again after, this is not the case
>>>>>>>> of the key length and the type of hash.
>>>>>>>>
>>>>>>>> There is no need to read the RSS config from the NIC, let's just
>>>>>>>> try to set the user defined one.
>>>>>>>>
>>>>>>>> Fixes: 8205e241b2b0 ("app/testpmd: add missing type to RSS hash
>>>>>>>> commands")
>>>>>>>> Cc: stable at dpdk.org
>>>>>>>> Cc: nelio.laranjeiro at 6wind.com
>>>>>>>>
>>>>>>>> Signed-off-by: Maxime Coquelin <maxime.coquelin at redhat.com>
>>>>>>>> ---
>>>>>>>>  app/test-pmd/config.c | 8 ++------
>>>>>>>>  1 file changed, 2 insertions(+), 6 deletions(-)
>>>>>>>>
>>>>>>>> diff --git a/app/test-pmd/config.c b/app/test-pmd/config.c index
>>>>>>>> 31d8ba1b91..451bda53b1 100644
>>>>>>>> --- a/app/test-pmd/config.c
>>>>>>>> +++ b/app/test-pmd/config.c
>>>>>>>> @@ -2853,18 +2853,14 @@ port_rss_hash_key_update(portid_t
>> port_id, char rss_type[], uint8_t *hash_key,
>>>>>>>>  	int diag;
>>>>>>>>  	unsigned int i;
>>>>>>>>
>>>>>>>> -	rss_conf.rss_key = NULL;
>>>>>>>> +	rss_conf.rss_key = hash_key;
>>>>>>>>  	rss_conf.rss_key_len = hash_key_len;
>>>>>>>>  	rss_conf.rss_hf = 0;
>>>>>>>>  	for (i = 0; rss_type_table[i].str; i++) {
>>>>>>>>  		if (!strcmp(rss_type_table[i].str, rss_type))
>>>>>>>>  			rss_conf.rss_hf = rss_type_table[i].rss_type;
>>>>>>>>  	}
>>>>>>>> -	diag = rte_eth_dev_rss_hash_conf_get(port_id, &rss_conf);
>>>>>>>> -	if (diag == 0) {
>>>>>>>> -		rss_conf.rss_key = hash_key;
>>>>>>>> -		diag = rte_eth_dev_rss_hash_update(port_id,
>> &rss_conf);
>>>>>>>> -	}
>>>>>>>> +	diag = rte_eth_dev_rss_hash_update(port_id, &rss_conf);
>>>>>>>
>>>>>>> I'm not 100% sure, but I'd say the intent above could be to
>>>>>>> update key only as the function name says. I.e. keep rss_hf as
>>>>>>> is. That could be the reason to get first.
> 
> +1
> The intent is only updating rss key. That's why to get_rss_conf first.
> At least for all intel devices. RSS key is a global value for all rss_hf.
> 
> And since the intent is to only update the key value. I don't think this patch makes sense since you're changing rss_hf which will break current test cases.
> And before 8205e241b2b01c, the command is what we want. port_rss_hash_key_update(portid_t port_id, uint8_t *hash_key) only updates key.
> 
> But if mlx has the configuration of config key for each rss_type then the code should remain to the current code in which keylen and rss_hf will be overridden to the correct value intel wants and mlx has their own configuration for specific rss type.
> But to be honest, I checked mlx5. I don't see they have this kind of configuration. Need to confirm with their driver maintainer though.
> 
> But anyway, from my point of view, I prefer to revert what 8205e241b2b01c0 does and remove rss_hf, rss_key_len setting in this command if mlx doesn't have key update for specific rss type. Otherwise, remain what it's like wight now.

For the main branch, we could revert 8205e241b2b01c0, but in my opinion,
we need to keep the hash_key_length otherwise it could lead to out of
bounds accesses if the key passed by the user is shorter than the key
length in use by the driver.

Note that it would also imply changes in the DTS and CIs tests that make
use of this command. We would also need to introduce a new command to be
able to set the rss hash types, and rename the existing one to be key-
specific. Otherwise we have no way with testpmd to configure RSS
properly. Given there does not seem to have any driver that leverages
RSS HF/Key pairs, maybe the simple thing is to just do what I did in
this patch.

For stable, I don't think this is a good idea to change testpmd
commands, as it could break users setups.

Thanks,
Maxime

> 
> BRs
> Xiaoyun
> 
>>>>
>>>> True,
>>>>
>>>>>> I think that was the intial purpose of the command, but patch
>>>>>> 8205e241b2b0 added setting the hash type as mandatory. There are
>>>>>> no other command to configure the hash type from testpmd AFAICT.
>>>>
>>>> Also for the same initial purpose, some NIC have an hash key per
>>>> protocol, by default it uses the same key for all of them but it can
>>>> be configured individually making for example key0 for all protocols
>>>> expect
>>>> IPv4 which uses key1.
>>>
>>> Thanks for the info, I have looked at most drivers but didn't found
>>> one that support this feature, could you give some pointer?
>>
>> Well, I have done the modification at that time for MLX5 PMD, since I left DPDK
>> in 2018 I don't know if they still support such configuration from this API or if
>> they fully moved to rte_flow.
>>
>>> Given how the drivers implément the callback, do you agree with the
>>> fix, or do you have something else in mind?
>>
>> I cannot answer if this get() is mandatory, this predates my arrival on DPDK
>> (original commit written in 2014), looking at DPDK state on  commit
>> f79959ea1504 ("app/testpmd: allow to configure RSS hash key").
>> Maybe someone from Intel can help, eventually you can contact PMD
>> maintainers to review this patch.
>>
>> Regards,
>> Nélio
>>
>>> Thanks,
>>> Maxime
>>>
>>>>>> Also, even without 8205e241b2b0, the function was broken because
>>>>>> the key length was overiden.
>>>>>
>>>>> I see, many thanks for explanations.
>>>>
>>>
>>
>> --
>> Nélio Laranjeiro
>> 6WIND
>