Re: [dpdk-dev] [PATCH v5] app/testpmd: fix the default RSS key configuration

[Ajit Khaparde <ajit.khaparde@broadcom.com>]

On Thu, Oct 15, 2020 at 4:54 PM Ferruh Yigit <ferruh.yigit@intel.com> wrote:
>
> On 10/16/2020 12:21 AM, Ophir Munk wrote:
> >> From: Ferruh Yigit <ferruh.yigit@intel.com>
> >>>>> On 10/15/2020 1:41 PM, Lijun Ou wrote:
> >>>>>> When start the testpmd, the pmd driver initializes the RSS
> >>>>>> configuration. Generally, the recommended RSS hash key is used as
> >>>>>> the default key in the driver. In addition, the default key is
> >>>>>> different from the default RSS flow in testpmd without specifying
> >>>>>> RSS hash key. So. if you do not specify the RSS key when creating
> >>>>>> an RSS rule, the testpmd uses the default key as the default RSS
> >>>>>> key of the RSS rule. As a result, you may mistakenly consider that
> >>>>>> the RSS key in use is the valid default key of the NIC, actually,
> >>>>>> the key and the valid default key of the NIC are two values.
> >>>>>>
> >>>>>> Consider the follow usage with testpmd:
> >>>>>> 1. first, startup testpmd:
> >>>>>> testpmd> show port 0 rss-hash key
> >>>>>> RSS functions:
> >>>>>>     all ipv4-frag ipv4-other ipv6-frag ipv6-other ip RSS key:
> >>>>>> -
> >> 6D5A56DA255B0EC24167253D43A38FB0D0CA2BCBAE7B30B477CB2DA3803
> >> 0F
> >>>>>> -20C6A42B73BBEAC01FA
> >>>>>> 2. create a rss rule
> >>>>>> testpmd> flow create 0 ingress pattern eth / ipv4 / udp / end
> >>>>>> testpmd> actions rss \
> >>>>>> types ipv4-udp end queues end / end
> >>>>>>
> >>>>>> 3. show rss-hash key
> >>>>>> testpmd> show port 0 rss-hash key
> >>>>>> RSS functions:
> >>>>>>     all ipv4-udp udp
> >>>>>> RSS key:
> >>>>>> -
> >> 74657374706D6427732064656661756C74205253532068617368206B65792C
> >> 206F
> >>>>>> -76657272696465
> >>>>>>
> >>>>>> In order to solve the above problems, it use the NIC valid default
> >>>>>> RSS key instead of the testpmd dummy RSS key in the flow
> >>>>>> configuration when the RSS key is not specified in the flow rule.
> >>>>>> If the NIC RSS key is invalid, it will use testpmd dummy RSS key as the
> >> default key.
> >>>>>>
> >>>>>> Fixes: ac8d22de2394 ("ethdev: flatten RSS configuration in flow
> >>>>>> API")
> >>>>>> Cc: stable@dpdk.org
> >>>>>>
> >>>>>> Signed-off-by: Lijun Ou <oulijun@huawei.com>
> >>>>>> Reviewed-by: Phil Yang <phil.yang@arm.com>
> >>>>>> ---
> >>>>>> V4->V5:
> >>>>>> -rewrite the commit log
> >>>>>> -add reviewed-by
> >>>>>
> >>>>> Hi Lijun,
> >>>>>
> >>>>> There were multiple other comments, it seems they are not addressed
> >>>>> but only updated the commit log, can you please check comments to
> >> prev versions.
> >>>>>
> >>>>> Before going into the details, my question was what happens if
> >>>>> default key not provided at all?
> >>>>> It seems this has been already tried by Ophir [1], later reverted
> >>>>> back [2] bringing the initial issue back.
> >>>>>
> >>>>> According commit, the reason of revert is to support following
> >> command:
> >>>>> "flow create 0 <pattern> actions rss queues 0 1 end key_len 40 / end"
> >>>>>
> >>>>> @Ophir, @Lijun,
> >>>>> Can we ignore the 'key_len' if the 'key' is not supported and solve
> >>>>> current issue as initially intended ([1])?
> >>>>>
> >>>> Hi, Ferruh
> >>>>     I have discussed with Phil Yang about the problem in [1]. I think
> >>>> there may be other problems with the idea and there is no better
> >>>> solution. and we need to remove key_len definition from rte_rss_conf
> >>>> structure. They don't have a plan. And [1] was eventually reverted.
> >>>>
> >>>
> >>> Why ignoring 'key_len' (set it to zero) when there is no 'key'
> >>> provided doesn't work?
> >>>
> >>
> >> What do you think [1] + following update, will it work?
> >>
> >>    diff --git a/lib/librte_ethdev/rte_flow.c b/lib/librte_ethdev/rte_flow.c
> >>    index ee4f3464fe..e7789c87b3 100644
> >>    --- a/lib/librte_ethdev/rte_flow.c
> >>    +++ b/lib/librte_ethdev/rte_flow.c
> >>    @@ -535,7 +535,7 @@ rte_flow_conv_action_conf(void *buf, const size_t
> >> size,
> >>                               }),
> >>                               size > sizeof(*dst.rss) ? sizeof(*dst.rss) : size);
> >>                    off = sizeof(*dst.rss);
> >>    -               if (src.rss->key_len) {
> >>    +               if (src.rss->key_len && src.rss->key) {
> >>                            off = RTE_ALIGN_CEIL(off, sizeof(*dst.rss->key));
> >>                            tmp = sizeof(*src.rss->key) * src.rss->key_len;
> >>                            if (size >= off + tmp)
> >>
> >
> > Ferruh, your suggestion ([1] + update) looks correct. I also verified it on mlx5 PMD.
> > Advantage: it's a generic fix for all dpdk applications using rte_flows (not just testpmd).
> > It reduces code.
> > With this fix the responsibility of handling key==NULL and/or len==0 is moved to the PMDs (which is good).
> >
> > With regard to Lijun patch - I liked the approach of overriding the default testpmd key with the default PMD key.
> > But it only addresses testpmd. More code was added.
> > It seems OK to call rte_eth_dev_rss_hash_conf_get() as part of parsing RSS, but it would feel more confident if we could confirm it for all the PMDs (by testing) or at least review the PMDs rss_hash_conf_get() implementations.
> >
>
> Lijun's idea can work. There was a problem in implementation related to the key
> size assumption, which can be fixed.
>
> Even it is fixed, when user gives a rss rule without a key, we are getting key
> from device and feeding same key back to device, this is unnecessary I think.
> When user didn't provide a key, rss rule shouldn't touch the key at all.
+1
But can we add this as expected behavior in the rte_flow document.

>
> Complication was when user provides key_len without a key, I think both ignoring
> or returning error in this case is OK.
Let's return an error if key_len is provided without a key.