Hi,
dpdk version in use is 19.11
I have not tried with latest upstream version.
It seems performance is affected by IPv6 neighbor advertisement packets coming
to this interface
05:20:04.025290 IP6 fe80::6cf1:9fff:fe4e:8a01 > ff02::1: ICMP6, neighbor
advertisement, tgt is fe80::6cf1:9fff:fe4e:8a01, length 32
0x0000: 3333 0000 0001 6ef1 9f4e 8a01 86dd 6008
0x0010: fe44 0020 3aff fe80 0000 0000 0000 6cf1
0x0020: 9fff fe4e 8a01 ff02 0000 0000 0000 0000
0x0030: 0000 0000 0001 8800 96d9 2000 0000 fe80
0x0040: 0000 0000 0000 6cf1 9fff fe4e 8a01 0201
0x0050: 6ef1 9f4e 8a01
Somehow, there are about 100 such packets per second coming to the interface,
and packet loss happens.
When we change default vlan in switch so that there is no such packets come to
the interface (the mlx5 VF under test), there is not packet loss anymore.
In both cases, all packets have arrived to rx_vport_unicast_packets.
In the packet loss case, we see less packets in rx_good_packets
(rx_vport_unicast_packets = rx_good_packets + lost packet).
If the dpdk application is too slow to receive all packets from the VF, is
there any counter to indicate this?
Any suggestion?
Thank you.
Best regards
Yan Xiaoping
-----Original Message-----
From: Asaf Penso <as...@nvidia.com>
Sent: 2021年7月13日 20:36
To: Yan, Xiaoping (NSB - CN/Hangzhou) <xiaoping....@nokia-sbell.com>;
users@dpdk.org
Cc: Slava Ovsiienko <viachesl...@nvidia.com>; Matan Azrad <ma...@nvidia.com>;
Raslan Darawsheh <rasl...@nvidia.com>
Subject: RE: mlx5 VF packet lost between rx_port_unicast_packets and
rx_good_packets
Hello Yan,
Can you please mention which DPDK version you use and whether you see this
issue also with latest upstream version?
Regards,
Asaf Penso
>-----Original Message-----
>From: users <users-boun...@dpdk.org<mailto:users-boun...@dpdk.org>> On Behalf
>Of Yan, Xiaoping (NSB -
>CN/Hangzhou)
>Sent: Monday, July 5, 2021 1:08 PM
>To: users@dpdk.org<mailto:users@dpdk.org>
>Subject: [dpdk-users] mlx5 VF packet lost between
>rx_port_unicast_packets and rx_good_packets
>
>Hi,
>
>When doing traffic loopback test on a mlx5 VF, we found there are some
>packet loss (not all packet received back ).
>
>From xstats counters, I found all packets have been received in
>rx_port_unicast_packets, but rx_good_packets has lower counter, and
>rx_port_unicast_packets - rx_good_packets = lost packets i.e. packet
>lost between rx_port_unicast_packets and rx_good_packets.
>But I can not find any other counter indicating where exactly those
>packets are lost.
>
>Any idea?
>
>Attached is the counter logs. (bf is before the test, af is after the
>test, fp-cli dpdk-port-stats is the command used to get xstats, and
>ethtool -S _f1 (the vf
>used) also printed) Test equipment reports that it sends: 2911176
>packets,
>receives: 2909474, dropped: 1702 And the xstats (after - before) shows
>rx_port_unicast_packets 2911177, rx_good_packets 2909475, so drop
>(2911177 - rx_good_packets) is 1702
>
>BTW, I also noticed this discussion "packet loss between phy and good
>counter"
>http://mails.dpdk.org/archives/users/2018-July/003271.html
>but my case seems to be different as packet also received in
>rx_port_unicast_packets, and I checked counter from pf (ethtool -S
>ens1f0 in attached log), rx_discards_phy is not increasing.
>
>Thank you.
>
>Best regards
>Yan Xiaoping
