[dpdk-dev] [PATCH v2] net/i40e: fix Rx statistic inconsistent

[Kyle Larose]

Hello Wei,


On Tue, Aug 2, 2016 at 2:59 AM, Zhao1, Wei <wei.zhao1 at intel.com> wrote:
> Hi, Wujingjing and Kyle Larose
>
>
>
>> -----Original Message-----
>> From: Zhao1, Wei
>> Sent: Tuesday, August 2, 2016 11:27 AM
>> To: Wu, Jingjing <jingjing.wu at intel.com>; Lu, Wenzhuo
>> <wenzhuo.lu at intel.com>
>> Cc: dev at dpdk.org
>> Subject: RE: [dpdk-dev] [PATCH v2] net/i40e: fix Rx statistic inconsistent
>>
>> Hi,Wu jingjing and wenzhuo
>>
>> > -----Original Message-----
>> > From: Zhao1, Wei
>> > Sent: Monday, August 1, 2016 4:58 PM
>> > To: 'Kyle Larose' <eomereadig at gmail.com>
>> > Cc: dev at dpdk.org
>> > Subject: RE: [dpdk-dev] [PATCH v2] net/i40e: fix Rx statistic
>> > inconsistent
>> >
>> > Hi, Kyle Larose
>> >    The core problem is i40e has no statistic of discard bytes, that
>> > means even if when ports are not stopped, the statistic  rx_good_bytes
>> > is consist of discard
>> > bytes，is that reasonable? In other words, I can just minus discard
>> > bytes from rx_good_bytes if I can get discard bytes number, that is much
>> better.
>> >
>> > -----Original Message-----
>> > From: Kyle Larose [mailto:eomereadig at gmail.com]
>> > Sent: Saturday, July 30, 2016 1:17 AM
>> > To: Zhao1, Wei <wei.zhao1 at intel.com>
>> > Cc: dev at dpdk.org
>> > Subject: Re: [dpdk-dev] [PATCH v2] net/i40e: fix Rx statistic
>> > inconsistent
>> >
>> > On Fri, Jul 29, 2016 at 4:50 AM, Wei Zhao1 <wei.zhao1 at intel.com> wrote:
>> > > rx_good_bytes and rx_good_packets statistic is inconsistent when
>> > > port stopped,ipackets statistic is minus the discard packets but
>> > > rx_bytes statistic not.Also,i40e has no statistic of discard bytes,
>> > > so we have to delete discard packets item from rx_good_packets statistic.
>> > >
>> > > Fixes: 9aace75fc82e ("i40e: fix statistics")
>> > >
>> > > Signed-off-by: Wei Zhao1 <wei.zhao1 at intel.com>
>> > > ---
>> > >  drivers/net/i40e/i40e_ethdev.c | 3 +--
>> > >  1 file changed, 1 insertion(+), 2 deletions(-)
>> > >
>> > > diff --git a/drivers/net/i40e/i40e_ethdev.c
>> > > b/drivers/net/i40e/i40e_ethdev.c index 11a5804..553dfd9 100644
>> > > --- a/drivers/net/i40e/i40e_ethdev.c
>> > > +++ b/drivers/net/i40e/i40e_ethdev.c
>> > > @@ -2319,8 +2319,7 @@ i40e_dev_stats_get(struct rte_eth_dev *dev,
>> > > struct rte_eth_stats *stats)
>> > >
>> > >         stats->ipackets = pf->main_vsi->eth_stats.rx_unicast +
>> > >                         pf->main_vsi->eth_stats.rx_multicast +
>> > > -                       pf->main_vsi->eth_stats.rx_broadcast -
>> > > -                       pf->main_vsi->eth_stats.rx_discards;
>> > > +                       pf->main_vsi->eth_stats.rx_broadcast;
>> > >         stats->opackets = pf->main_vsi->eth_stats.tx_unicast +
>> > >                         pf->main_vsi->eth_stats.tx_multicast +
>> > >                         pf->main_vsi->eth_stats.tx_broadcast;
>> > > --
>> > > 2.5.5
>> > >
>> >
>> > Is it not worse to report a received packet when no packet was
>> > actually received by the upper layers under normal operations than to
>> > ensure that packets and  bytes are consistent when an interface is
>> > stopped? It seems like the first case is much more likely to occur than the
>> second.
>> >
>> > Are we just introducing a new issue to fix another?
>> >
>> > How does this behaviour compare to other NICs? Does the ixgbe report
>> > discarded packets in its ipackets? My reading of the driver is that it does not.
>> > In fact, it does something interesting to deal with the
>> > problem:
>> >
>> > from:
>> > http://dpdk.org/browse/dpdk/tree/drivers/net/ixgbe/ixgbe_ethdev.c
>> >
>> > /*
>> > * An errata states that gprc actually counts good + missed packets:
>> > * Workaround to set gprc to summated queue packet receives */
>> > hw_stats-
>> > >gprc = *total_qprc;
>> >
>> > total_gprc is equal to the sum of the qprc per queue. Can we do
>> > something similar on the i40e instead of adding unicast, mulitcast and
>> broadcast?
>>
>>
>> I have checked ixgbe code about  Rx statistic, in function
>> ixgbe_read_stats_registers() we can find the rx_good_bytes and
>> rx_good_packets statistic.
>> It is listed below, we  can see rx_good_packets is also just addition of Queue
>> Packets Received Count and  not minused  discard packet number.
>> Is there some wrong of understanding?



My understanding of the problem can be broken into three parts:
 1) In Unicast/Multicast/Broadcast packet counters are counting
packets which were discarded
 2) The corresponding byte counters count packets which were discarded.
 3) There are no discarded byte counters.

Our in bytes counter consists of the sum of in unicast, in multicast,
and in broadcast. This sum includes discarded bytes, which we do not
want, for two reasons. First, it would lead to misleading bitrate
reports: people expect to see the amount of traffic actually handled.
Second, it conflicts with the current packet counters (the counters
without your change).

Obviously if we could count the discarded bytes, we could subtract
them. Alternatively, if we could count only received bytes which were
not discarded, then we would not need to subtract discarded bytes from
the received bytes counter. The ixgbe had a similar problem for
packets, but presumably not bytes. So, I looked to it for inspiration.

Presumably, the ixgbe per-queue counters count all received packets,
per queue, and do not count discarded packets.Thus, is you take the
sum of received packets across all queues, you can find the total
number of received packets without discards included.

Can we not do the same thing on the i40e, but for bytes? I just took a
quick look at the datasheet, and I cannot see anything offhand that is
equivalent to qbrc, which is what I would have used here. That said,
I'm not an i40e expert, so I may have missed something. That's why I'm
asking if it's possible. :)



>>
>> for (i = 0; i < IXGBE_QUEUE_STAT_COUNTERS; i++) {
>> ......
>>                                  *total_qprc += hw_stats->qprc[i];
>>               *total_qbrc += hw_stats->qbrc[i];
>> ......
>> }
>
>    The problem is i40e has no statistic of discard bytes, so it is impossible to minus discard bytes from rx_good_bytes . If you think it's not reasonable to
> Delete  rx_discards iterm from rx_good_packets statistic, this patch will be superseded. Because I didn't find other way to correct this problem at present.
>

I understand that there is no discard byte counter. Are there
per-queue byte counters on the i40e, which do not count discarded
byte, with which we could do a similar thing as the packet counters on
the ixgbe?

My concern is that both solutions aside from that are not good: either
we count less packets received than bytes received, or we count
discarded packets as received packets. Is there any chance that a
firmware update in the future could fix this? This sort of
inconsistency, or inability to provide *good* stats, really sucks for
a NIC.

Thanks,

Kyle