[dpdk-users] Mellanox MLX5 - RX Missed Errors Counter Overflow

[Stephen Hemminger]

On Tue, 14 Jul 2020 10:35:25 +0200
Tomáš Jánský <tomas.jansky at flowmon.com> wrote:

> Hi,
> 
> I am using the Mellanox MLX5 driver and I noticed nonsense values in
> *rx_missed_errors* counter.
> 
> Configuration:
> 
>    - DPDK version 20.02
>    - CentOS Linux release 7.7.1908 (Core)
>    - Mellanox OFED version: MLNX_OFED_LINUX-4.7-3.2.9.0-rhel7.7-x86_64
> 
> 
> I have a function which periodically retrieves xstats counters.
> 
> char *field_names[7] = {"rx_packets_phy", "rx_discards_phy",
> "rx_missed_errors", "rx_errors", "rx_bytes_phy", "tx_packets_phy",
> "tx_bytes_phy"};
> uint64_t ids[7], values[7];
> int i;
> 
> for (i = 0; i < 7; i++) {
>     rte_eth_xstats_get_id_by_name(port_id, field_names[i], &ids[i]);
> }
> rte_eth_xstats_get_by_id(port_id, ids, values, 7);
> 
> It works perfectly fine until the value in *rx_missed_errors* counter
> reaches around 3721692877 (maximum observed value, but probably not the
> actual maximum value) which is 0xDDD486CD hex.
> After that, the next measurement reports the value 18446744073150886746
> which is obviously wrong.
> Then, after some time, the value starts again from zero.
> 
> The documentation says that the counter is supposed to be 64-byte
> unsigned integer: https://doc.dpdk.org/api/structrte__eth__xstat.html
> 
> Is anybody else experiencing this issue? It might not be related only to
> the MLX5 driver because we had a similar issue with Intel X710 cards.
> 
> Thanks.
> 
> Tomas

Xstats are often created by the driver from hardware registers.
Investigate which hardware register that counter is coming from and see
what the width of the register is.

You may have to use the 32 bit counter detection logic (like snmpd)
to workaround.