Re: [dpdk-dev] [PATCH] app/testpmd: fix doubling of 'total TX dropped'

[David Marchand <david.marchand@redhat.com>]

On Tue, Jul 16, 2019 at 5:29 PM Ferruh Yigit <ferruh.yigit@intel.com> wrote:
>
> On 7/16/2019 4:00 PM, Andrew Rybchenko wrote:
> > On 7/16/19 5:28 PM, Bruce Richardson wrote:
> >> On Tue, Jul 16, 2019 at 03:23:03PM +0100, Ferruh Yigit wrote:
> >>> On 7/16/2019 1:23 PM, Andrew Rybchenko wrote:
> >>>> On 7/15/19 5:53 PM, Ferruh Yigit wrote:
> >>>>> On 7/12/2019 9:32 AM, A.McLoughlin wrote:
> >>>>>> The 'Accumulated forward statistics for all ports' incorrectly displayed
> >>>>>> double the actual value for 'total_tx_dropped'. This was because 2
> >>>>>> lines in the same function both incremented total_tx_dropped every time
> >>>>>> a packet was dropped.  I removed one of these lines to fix this issue.
> >>>>>>
> >>>>>> Fixes: 53324971a14e ("app/testpmd: display/clear forwarding stats on demand")
> >>>>>> Cc: david.marchand@redhat.com
> >>>>>> Cc: stable@dpdk.org
> >>>>>>
> >>>>>> Signed-off-by: A.McLoughlin <aideen.mcloughlin@intel.com>
> >>>>>> ---
> >>>>>>    app/test-pmd/testpmd.c | 1 -
> >>>>>>    1 file changed, 1 deletion(-)
> >>>>>>
> >>>>>> diff --git a/app/test-pmd/testpmd.c b/app/test-pmd/testpmd.c
> >>>>>> index 3ed3523b7..c41bada50 100644
> >>>>>> --- a/app/test-pmd/testpmd.c
> >>>>>> +++ b/app/test-pmd/testpmd.c
> >>>>>> @@ -1555,7 +1555,6 @@ fwd_stats_display(void)
> >>>>>>                  total_recv += stats.ipackets;
> >>>>>>                  total_xmit += stats.opackets;
> >>>>>>                  total_rx_dropped += stats.imissed;
> >>>>>> -                total_tx_dropped += ports_stats[pt_id].tx_dropped;
> >>>>>>                  total_tx_dropped += stats.oerrors;
> >>>>>>                  total_rx_nombuf  += stats.rx_nombuf;
> >>>>>>
> >>>>>>
> >>>>>
> >>>>> Hi Aideen,
> >>>>>
> >>>>> Indeed 'rte_eth_stats->oerrors' and 'tx_dropped' are different values,
> >>>>>
> >>>>> in testpmd, 'TX-total' is taken as "total_xmit + total_tx_dropped", from this
> >>>>> description it may be fair to say
> >>>>> "total_tx_dropped = oerrors + tx_dropped"
> >>>>>
> >>>>> This is easier to see in HW devices, 'oerrors' is the packets sent to HW but HW
> >>>>> reported failure for some reason, so these packets not transmitted to the medium.
> >>>>> 'tx_dropped' is mostly calculated by application, driver returns packets that
> >>>>> can't able to sent to HW, so application can re-try to send or free them and
> >>>>> increase 'tx_dropped' counter.
> >>>>>
> >>>>>
> >>>>> The problem is in the virtual drivers, the packets not able to sent are
> >>>>> calculated as 'oerrors' and tx_burst functions returns the number of the
> >>>>> successfully sent packets which cause application calculate remaining ones as
> >>>>> 'tx_dropped' which cause the duplication.
> >>>>
> >>>> I don't understand how it is. Tx burst returns a number of owned packets
> >>>> (either successfully transmitted or internally dropped/freed). If it is
> >>>> smaller than number of packets in request, other packets are either
> >>>> retried or calculated as tx_dropped.
> >>>
> >>> Virtual PMDs, at least the ones I checked, calculating not sent packets as
> >>> error, also application calculates them as tx_dropped.
> >>>
> >>> Like:
> >>> https://git.dpdk.org/dpdk/tree/drivers/net/ring/rte_eth_ring.c?h=v19.08-rc1#n97
> >>>
> >> That is probably incorrect to do. Virtual PMDs should behave as real ones
> >> do as far as possible. I think we should change them to not count as errors
> >> any that are not handled by the driver, provided those are returned to the
> >> app.
> >
> > I agree with Bruce. It is a bug in ring PMD. If so, too fast attempts to
> > transmit packets will blow up err_pkts counter.
> >
>
> +1, as far as I can see following PMDs requires fixing:
> ring
> kni
> pcap
> tap

I already have fixes for some of them (found when I started to look at
the stats).


-- 
David Marchand