Re: [PATCH net-next v25 08/13] rtase: Implement net_device_ops

[Joe Damato <jdamato@fastly.com>]

On Mon, Jul 29, 2024 at 11:39:03AM +0000, Justin Lai wrote:
> > On Mon, Jul 29, 2024 at 02:21:16PM +0800, Justin Lai wrote:
> > > 1. Implement .ndo_set_rx_mode so that the device can change address
> > > list filtering.
> > > 2. Implement .ndo_set_mac_address so that mac address can be changed.
> > > 3. Implement .ndo_change_mtu so that mtu can be changed.
> > > 4. Implement .ndo_tx_timeout to perform related processing when the
> > > transmitter does not make any progress.
> > > 5. Implement .ndo_get_stats64 to provide statistics that are called
> > > when the user wants to get network device usage.
> > > 6. Implement .ndo_vlan_rx_add_vid to register VLAN ID when the device
> > > supports VLAN filtering.
> > > 7. Implement .ndo_vlan_rx_kill_vid to unregister VLAN ID when the
> > > device supports VLAN filtering.
> > > 8. Implement the .ndo_setup_tc to enable setting any "tc" scheduler,
> > > classifier or action on dev.
> > > 9. Implement .ndo_fix_features enables adjusting requested feature
> > > flags based on device-specific constraints.
> > > 10. Implement .ndo_set_features enables updating device configuration
> > > to new features.
> > >
> > > Signed-off-by: Justin Lai <justinlai0215@realtek.com>
> > > ---
> > >  .../net/ethernet/realtek/rtase/rtase_main.c   | 235 ++++++++++++++++++
> > >  1 file changed, 235 insertions(+)
> > >
> > > diff --git a/drivers/net/ethernet/realtek/rtase/rtase_main.c
> > > b/drivers/net/ethernet/realtek/rtase/rtase_main.c
> > > index 8fd69d96219f..80673fa1e9a3 100644
> > 
> > [...]
> > 
> > > +static void rtase_dump_state(const struct net_device *dev) {
> > 
> > [...]
> > 
> > > +
> > > +     netdev_err(dev, "tx_packets %lld\n",
> > > +                le64_to_cpu(counters->tx_packets));
> > > +     netdev_err(dev, "rx_packets %lld\n",
> > > +                le64_to_cpu(counters->rx_packets));
> > > +     netdev_err(dev, "tx_errors %lld\n",
> > > +                le64_to_cpu(counters->tx_errors));
> > > +     netdev_err(dev, "rx_errors %d\n",
> > > +                le32_to_cpu(counters->rx_errors));
> > > +     netdev_err(dev, "rx_missed %d\n",
> > > +                le16_to_cpu(counters->rx_missed));
> > > +     netdev_err(dev, "align_errors %d\n",
> > > +                le16_to_cpu(counters->align_errors));
> > > +     netdev_err(dev, "tx_one_collision %d\n",
> > > +                le32_to_cpu(counters->tx_one_collision));
> > > +     netdev_err(dev, "tx_multi_collision %d\n",
> > > +                le32_to_cpu(counters->tx_multi_collision));
> > > +     netdev_err(dev, "rx_unicast %lld\n",
> > > +                le64_to_cpu(counters->rx_unicast));
> > > +     netdev_err(dev, "rx_broadcast %lld\n",
> > > +                le64_to_cpu(counters->rx_broadcast));
> > > +     netdev_err(dev, "rx_multicast %d\n",
> > > +                le32_to_cpu(counters->rx_multicast));
> > > +     netdev_err(dev, "tx_aborted %d\n",
> > > +                le16_to_cpu(counters->tx_aborted));
> > > +     netdev_err(dev, "tx_underun %d\n",
> > > +                le16_to_cpu(counters->tx_underun));
> > 
> > You use le64/32/16_to_cpu here for all stats, but below in rtase_get_stats64, it
> > is only used for tx_errors.
> > 
> > The code should probably be consistent? Either you do or don't need to use
> > them?
> > 
> > > +}
> > > +
> > [...]
> > > +
> > > +static void rtase_get_stats64(struct net_device *dev,
> > > +                           struct rtnl_link_stats64 *stats) {
> > > +     const struct rtase_private *tp = netdev_priv(dev);
> > > +     const struct rtase_counters *counters;
> > > +
> > > +     counters = tp->tally_vaddr;
> > > +
> > > +     dev_fetch_sw_netstats(stats, dev->tstats);
> > > +
> > > +     /* fetch additional counter values missing in stats collected by driver
> > > +      * from tally counter
> > > +      */
> > > +     rtase_dump_tally_counter(tp);
> > > +     stats->rx_errors = tp->stats.rx_errors;
> > > +     stats->tx_errors = le64_to_cpu(counters->tx_errors);
> > > +     stats->rx_dropped = tp->stats.rx_dropped;
> > > +     stats->tx_dropped = tp->stats.tx_dropped;
> > > +     stats->multicast = tp->stats.multicast;
> > > +     stats->rx_length_errors = tp->stats.rx_length_errors;
> > 
> > See above; le64_to_cpu for tx_errors, but not the rest of the stats. Why?
> 
> The rtase_dump_state() function is primarily used to dump certain hardware
> information. Following discussions with Jakub, it was suggested that we
> should design functions to accumulate the 16-bit and 32-bit counter values
> to prevent potential overflow issues due to the limited size of the
> counters. However, the final decision was to temporarily refrain from
> reporting 16-bit and 32-bit counter information. Additionally, since
> tx_packet and rx_packet data are already provided through tstat, we
> ultimately opted to modify it to the current rtase_get_stats64() function.

Your response was a bit confusing, but after re-reading the code I
think I understand now that I misread the code above.

The answer seems to be that tx_errors is accumulated in
rtase_counters (which needs le*_to_cpu), but the other counters are
accumulated in tp->stats which do not need le*_to_cpu because they
are already being accounted in whatever endianness the CPU uses.

OK.