On Thu, May 19, 2022 at 2:14 PM David Laight <David.Laight@aculab.com> wrote:
>
> From: Michael Chan
> > Sent: 19 May 2022 01:52
> >
> > On Wed, May 18, 2022 at 2:31 PM David Laight <David.Laight@aculab.com> wrote:
> > >
> > > From: Paolo Abeni
> > > > Sent: 18 May 2022 18:27
> > > ....
> > > > > If I read /sys/class/net/em2/statistics/rx_packets every second
> > > > > delaying with:
> > > > >   syscall(SYS_clock_nanosleep, CLOCK_MONOTONIC, TIMER_ABSTIME, &ts, NULL);
> > > > > about every 43 seconds I get a zero increment.
> > > > > This really doesn't help!
> > > >
> > > > It looks like the tg3 driver fetches the H/W stats once per second. I
> > > > guess that if you fetch them with the same period and you are unlucky
> > > > you can read the same sample 2 consecutive time.
> > >
> > > Actually I think the hardware is writing them to kernel memory
> > > every second.
> >
> > On your BCM95720 chip, statistics are gathered by tg3_timer() once a
> > second.  Older chips will use DMA.
>
> Ah, I wasn't sure which code was relevant.
> FWIW the code could rotate 64bit values by 32 bits
> to convert to/from the strange ordering the hardware uses.
>
> > Please show a snapshot of all the counters.  In particular,
> > rxbds_empty, rx_discards, etc will show whether the driver is keeping
> > up with incoming RX packets or not.
>
> After running the test for a short time.
> The application stats indicate that around 40000 packets are missing.
>
> # ethtool -S em2 | grep -v ' 0$'; for f in /sys/class/net/em2/statistics/*; do echo $f $(cat $f); done|grep -v ' 0$'
> NIC statistics:
>      rx_octets: 4589028558
>      rx_ucast_packets: 21049866
>      rx_mcast_packets: 763
>      rx_bcast_packets: 746
>      tx_octets: 4344
>      tx_ucast_packets: 6
>      tx_mcast_packets: 40
>      tx_bcast_packets: 3
>      rxbds_empty: 76
>      rx_discards: 14
>      mbuf_lwm_thresh_hit: 14
> /sys/class/net/em2/statistics/multicast 763
> /sys/class/net/em2/statistics/rx_bytes 4589028558
> /sys/class/net/em2/statistics/rx_missed_errors 14
> /sys/class/net/em2/statistics/rx_packets 21433169
> /sys/class/net/em2/statistics/tx_bytes 4344
> /sys/class/net/em2/statistics/tx_packets 49
>
> I've replaced the rx_packets count with an atomic64 counter in tg3_rx().
> Reading every second gives values like:
>
> # echo_every 1 |(c=0; n0=0; while read r; do n=$(cat /sys/class/net/em2/statistics/rx_packets); echo $c $((n - n0)); c=$((c+1)); n0=$n; done)
> 0 397169949
> 1 399831
> 2 399883
> 3 399913
> 4 399871
> 5 398747
> 6 400035
> 7 399958
> 8 399947
> 9 399923
> 10 399978
> 11 399457
> 12 399130
> 13 400128
> 14 399808
> 15 399029
>

I see that in a span of 15 seconds, the received packets are 4362 less
than what you are expecting (considering 400000/s avg)
In what time period did the application report 40000 missing packets?
Does it map to about 150 seconds of test time?
The error counters do not look suspicious at this point for the
reported problem.
Do you see this problem with any other traffic pattern?

> They should all be 400000 with slight variances.
> But there are clearly 100s of packets being discarded in some
> 1 second periods.
>
> I don't think I can blame the network.
> All the systems are plugged into the same ethernet switch on a test LAN.
>
>         David
>
> -
> Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
> Registration No: 1397386 (Wales)