Re: [PATCH] net/gve: add support for basic stats

[Rushil Gupta <rushilg@google.com>]

Hi all
Josh just found out some inconsistencies in the Tx/Rx statistics sum
for all ports. Not sure if we can screenshot here but it goes like
this:
Tx-dropped for port0: 447034
Tx-dropped for port1: 0
Accumulated forward statistics for all ports: 807630

Please note that this issue is only with Tx-dropped (not Tx-packets/Tx-total).


On Wed, Dec 7, 2022 at 8:39 AM Stephen Hemminger
<stephen@networkplumber.org> wrote:
>
> On Wed, 7 Dec 2022 15:09:08 +0000
> Ferruh Yigit <ferruh.yigit@amd.com> wrote:
>
> > On 11/24/2022 7:33 AM, Junfeng Guo wrote:
> > > Add support for dev_ops of stats_get and stats_reset.
> > >
> > > Queue stats update will be moved into xstat [1], but the basic stats
> > > items may still be required. So we just keep the remaining ones and
> > > will implement the queue stats via xstats in the coming release.
> > >
> > > [1]
> > > https://elixir.bootlin.com/dpdk/v22.07/ \
> > >     source/doc/guides/rel_notes/deprecation.rst#L118
> > >
> > > Signed-off-by: Xiaoyun Li <xiaoyun.li@intel.com>
> > > Signed-off-by: Junfeng Guo <junfeng.guo@intel.com>
> >
> > <...>
> >
> > > +static int
> > > +gve_dev_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
> > > +{
> > > +   uint16_t i;
> > > +
> > > +   for (i = 0; i < dev->data->nb_tx_queues; i++) {
> > > +           struct gve_tx_queue *txq = dev->data->tx_queues[i];
> > > +           if (txq == NULL)
> > > +                   continue;
> > > +
> > > +           stats->opackets += txq->packets;
> > > +           stats->obytes += txq->bytes;
> > > +           stats->oerrors += txq->errors;
> >
> > Hi Junfeng, Qi, Jingjing, Beilei,
> >
> > Above logic looks wrong to me, did you test it?
> >
> > If the 'gve_dev_stats_get()' called multiple times (without stats reset
> > in between), same values will be keep added to stats.
> > Some hw based implementations does this, because reading from stats
> > registers automatically reset those registers but this shouldn't be case
> > for this driver.
> >
> > I expect it to be something like:
> >
> > local_stats = 0
> > foreach queue
> >       local_stats += queue->stats
> > stats = local_stats
>
> The zero of local_stats is unnecessary.
> The only caller of the PMD stats_get is rte_ethdev_stats_get
> and it zeros the stats structure before calling the PMD.
>
>
> int
> rte_eth_stats_get(uint16_t port_id, struct rte_eth_stats *stats)
> {
>         struct rte_eth_dev *dev;
>
>         RTE_ETH_VALID_PORTID_OR_ERR_RET(port_id, -ENODEV);
>         dev = &rte_eth_devices[port_id];
>
>         memset(stats, 0, sizeof(*stats));
> ...
>         stats->rx_nombuf = dev->data->rx_mbuf_alloc_failed;
>         return eth_err(port_id, (*dev->dev_ops->stats_get)(dev, stats));
>
> If any PMD has extra memset in their stats get that could be removed.