[dpdk-dev] [External] Re: [PATCH v2] app/testpmd: flowgen support ip and udp fields

[王志宏]

On Tue, Aug 10, 2021 at 5:12 PM Ferruh Yigit <ferruh.yigit at intel.com> wrote:
>
> On 8/10/2021 8:57 AM, 王志宏 wrote:
> > Thanks for the review Ferruh :)
> >
> > On Mon, Aug 9, 2021 at 11:18 PM Ferruh Yigit <ferruh.yigit at intel.com> wrote:
> >>
> >> On 8/9/2021 7:52 AM, Zhihong Wang wrote:
> >>> This patch aims to:
> >>>  1. Add flexibility by supporting IP & UDP src/dst fields
> >>
> >> What is the reason/"use case" of this flexibility?
> >
> > The purpose is to emulate pkt generator behaviors.
> >
>
> 'flowgen' forwarding is already to emulate pkt generator, but it was only
> changing destination IP.
>
> What additional benefit does changing udp ports of the packets brings? What is
> your usecase for this change?

Pkt generators like pktgen/trex/ixia/spirent can change various fields
including ip/udp src/dst.

Keeping the cfg_n_* while setting cfg_n_ip_dst = 1024 and others = 1
makes the default behavior exactly unchanged. Do you think it makes
sense?

>
> >>
> >>>  2. Improve multi-core performance by using per-core vars>
> >>
> >> On multi core this also has syncronization problem, OK to make it per-core. Do
> >> you have any observed performance difference, if so how much is it?
> >
> > Huge difference, one example: 8 core flowgen -> rxonly results: 43
> > Mpps (per-core) vs. 9.3 Mpps (shared), of course the numbers "varies
> > depending on system configuration".
> >
>
> Thanks for clarification.
>
> >>
> >> And can you please separate this to its own patch? This can be before ip/udp update.
> >
> > Will do.
> >
> >>
> >>> v2: fix assigning ip header cksum
> >>>
> >>
> >> +1 to update, can you please make it as seperate patch?
> >
> > Sure.
> >
> >>
> >> So overall this can be a patchset with 4 patches:
> >> 1- Fix retry logic (nb_rx -> nb_pkt)
> >> 2- Use 'rte_ipv4_cksum()' API (instead of static 'ip_sum()')
> >> 3- User per-core varible (for 'next_flow')
> >> 4- Support ip/udp src/dst variaty of packets
> >>
> >
> > Great summary. Thanks a lot.
> >
> >>> Signed-off-by: Zhihong Wang <wangzhihong.wzh at bytedance.com>
> >>> ---
> >>>  app/test-pmd/flowgen.c | 137 +++++++++++++++++++++++++++++++------------------
> >>>  1 file changed, 86 insertions(+), 51 deletions(-)
> >>>
> >>
> >> <...>
> >>
> >>> @@ -185,30 +193,57 @@ pkt_burst_flow_gen(struct fwd_stream *fs)
> >>>               }
> >>>               pkts_burst[nb_pkt] = pkt;
> >>>
> >>> -             next_flow = (next_flow + 1) % cfg_n_flows;
> >>> +             if (++next_udp_dst < cfg_n_udp_dst)
> >>> +                     continue;
> >>> +             next_udp_dst = 0;
> >>> +             if (++next_udp_src < cfg_n_udp_src)
> >>> +                     continue;
> >>> +             next_udp_src = 0;
> >>> +             if (++next_ip_dst < cfg_n_ip_dst)
> >>> +                     continue;
> >>> +             next_ip_dst = 0;
> >>> +             if (++next_ip_src < cfg_n_ip_src)
> >>> +                     continue;
> >>> +             next_ip_src = 0;
> >>
> >> What is the logic here, can you please clarifiy the packet generation logic both
> >> in a comment here and in the commit log?
> >
> > It's round-robin field by field. Will add the comments.
> >
>
> Thanks. If the receiving end is doing RSS based on IP address, dst address will
> change in every 100 packets and src will change in every 10000 packets. This is
> a slight behavior change.
>
> When it was only dst ip, it was simple to just increment it, not sure about it
> in this case. I wonder if we should set all randomly for each packet. I don't
> know what is the better logic here, we can discuss it more in the next version.

A more sophisticated pkt generator provides various options among
"step-by-step" / "random" / etc.

But supporting multiple fields naturally brings this implicitly. It
won't be a problem as it can be configured by setting the cfg_n_* as
we discussed above.

I think rte_rand() is a good option, anyway this can be tweaked easily
once the framework becomes shaped.

>
> >>
> >>>       }
> >>>
> >>>       nb_tx = rte_eth_tx_burst(fs->tx_port, fs->tx_queue, pkts_burst, nb_pkt);
> >>>       /*
> >>>        * Retry if necessary
> >>>        */
> >>> -     if (unlikely(nb_tx < nb_rx) && fs->retry_enabled) {
> >>> +     if (unlikely(nb_tx < nb_pkt) && fs->retry_enabled) {
> >>>               retry = 0;
> >>> -             while (nb_tx < nb_rx && retry++ < burst_tx_retry_num) {
> >>> +             while (nb_tx < nb_pkt && retry++ < burst_tx_retry_num) {
> >>>                       rte_delay_us(burst_tx_delay_time);
> >>>                       nb_tx += rte_eth_tx_burst(fs->tx_port, fs->tx_queue,
> >>> -                                     &pkts_burst[nb_tx], nb_rx - nb_tx);
> >>> +                                     &pkts_burst[nb_tx], nb_pkt - nb_tx);
> >>>               }
> >>
> >> +1 to this fix, thanks for it. But can you please make a seperate patch for
> >> this, with proper 'Fixes:' tag etc..
> >
> > Ok.
> >
> >>
> >>>       }
> >>> -     fs->tx_packets += nb_tx;
> >>>
> >>>       inc_tx_burst_stats(fs, nb_tx);
> >>> -     if (unlikely(nb_tx < nb_pkt)) {
> >>> -             /* Back out the flow counter. */
> >>> -             next_flow -= (nb_pkt - nb_tx);
> >>> -             while (next_flow < 0)
> >>> -                     next_flow += cfg_n_flows;
> >>> +     fs->tx_packets += nb_tx;
> >>> +     /* Catch up flow idx by actual sent. */
> >>> +     for (i = 0; i < nb_tx; ++i) {
> >>> +             RTE_PER_LCORE(_next_udp_dst) = RTE_PER_LCORE(_next_udp_dst) + 1;
> >>> +             if (RTE_PER_LCORE(_next_udp_dst) < cfg_n_udp_dst)
> >>> +                     continue;
> >>> +             RTE_PER_LCORE(_next_udp_dst) = 0;
> >>> +             RTE_PER_LCORE(_next_udp_src) = RTE_PER_LCORE(_next_udp_src) + 1;
> >>> +             if (RTE_PER_LCORE(_next_udp_src) < cfg_n_udp_src)
> >>> +                     continue;
> >>> +             RTE_PER_LCORE(_next_udp_src) = 0;
> >>> +             RTE_PER_LCORE(_next_ip_dst) = RTE_PER_LCORE(_next_ip_dst) + 1;
> >>> +             if (RTE_PER_LCORE(_next_ip_dst) < cfg_n_ip_dst)
> >>> +                     continue;
> >>> +             RTE_PER_LCORE(_next_ip_dst) = 0;
> >>> +             RTE_PER_LCORE(_next_ip_src) = RTE_PER_LCORE(_next_ip_src) + 1;
> >>> +             if (RTE_PER_LCORE(_next_ip_src) < cfg_n_ip_src)
> >>> +                     continue;
> >>> +             RTE_PER_LCORE(_next_ip_src) = 0;
> >>> +     }
> >>
> >> Why per-core variables are not used in forward function, but local variables
> >> (like 'next_ip_src' etc..) used? Is it for the performance, if so what is the
> >> impact?
> >>
> >> And why not directly assign from local variables to per-core variables, but have
> >> above catch up loop?
> >>
> >>
> >
> > Local vars are for generating pkts, global ones catch up finally when
> > nb_tx is clear.
>
> Why you are not using global ones to generate packets? This removes the need for
> catch up?

When there are multiple fields, back out the overran index caused by
dropped packets is not that straightforward -- It's the "carry" issue
in adding.

>
> > So flow indexes only increase by actual sent pkt number.
> > It serves the same purpose of the original "/* backout the flow counter */".
> > My math isn't good enough to make it look more intelligent though.
> >
>
> Maybe I am missing something, for this case why not just assign back from locals
> to globals?

As above.

However, this can be simplified if we discard the "back out"
mechanism: generate 32 pkts and send 20 of them while the rest 12 are
dropped, the difference is that is the idx gonna start from 21 or 33
next time?