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

["Nélio Laranjeiro" <nelio.laranjeiro@6wind.com>]

+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@dpdk.org
> >>>>> Cc: nelio.laranjeiro@6wind.com
> >>>>>
> >>>>> Signed-off-by: Maxime Coquelin <maxime.coquelin@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.
> > 
> > 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