[ovs-dev] [PATCH v5 1/5] dpif-netdev: associate flow with a mark id

[Finn Christensen]

Hi Yuanhan,

This will not break our PMD. I think PMDs should be able to handle rss_conf == NULL and then failover to default or initially set rss_conf.

Finn

>-----Original Message-----
>From: Yuanhan Liu [mailto:yliu at fridaylinux.org]
>Sent: 29. januar 2018 07:59
>To: Stokes, Ian <ian.stokes at intel.com>
>Cc: dev at openvswitch.org; Finn Christensen <fc at napatech.com>; Darrell Ball
><dball at vmware.com>; Chandran, Sugesh <sugesh.chandran at intel.com>;
>Simon Horman <simon.horman at netronome.com>
>Subject: Re: [PATCH v5 1/5] dpif-netdev: associate flow with a mark id
>
>On Fri, Jan 26, 2018 at 04:19:30PM +0800, Yuanhan Liu wrote:
>> > > +static bool
>> > > +flow_mark_has_no_ref(uint32_t mark) {
>> > > +    struct dp_netdev_flow *flow;
>> > > +
>> >
>> > Maybe I'm missing something below, but I expected a hash to be
>computed for mark before being called with CMAP_FOR_EACH_WITH_HASH?
>>
>> I treat "mark" as the hash, as it's with the same type of "hash" and
>> it's uniq. But you are probably right, it might be better to get the
>> hash by "hash_int". Will fix it.
>
>Oops, I forgot to do the coressponding change for mark_to_flow find. Thus,
>the partial offload is completely skiped, as retrieving the flow from mark
>would always fail (returing NULL).
>
>The reason I missed it, while I was testing v6, is I found the flow creation is
>failed. I don't really change anything between v5 and v6 and when I was back
>to v5, I also met same issue. I then thought it might be introduced by the
>OFED or firmware change I have done (for switching to other projects).
>I then thought it's something I could figure out laterly. Thus, v6 was sent out
>basically just with a build test.
>
>I then figured out today that the flow creation failure is introduced by a MLX5
>PMD driver change in DPDK v17.11. It looks like a bug to me. And there are 2
>solutions for that:
>
>- fix it in MLX5 PMD (if it's a bug); I was talking to the author made
>  such change.
>- set rss_conf to NULL, which will let DPDK to follow the one OVS-DPDK
>  has set in the beginning.
>
>I chose 2, sicne option 1 won't change the fact that it won't work with DPDK
>v17.11.
>
>And Finn, I probably need your help to verify that option 2 won't break Napa
>PMD drivers.
>
>I will send v7 soon. Please help review.
>
>Thanks.
>
>	--yliu