Re: [dpdk-users] i40e VF Rx/Tx CRC Bytes Discrepancy

Re: [dpdk-users] i40e VF Rx/Tx CRC Bytes Discrepancy

[Stephen Hemminger]

On Thu, 8 Feb 2018 11:52:55 -0500
Matthew Coppola <matthew.coppola4@gmail.com> wrote:

> Hi,
> 
> I am noticing a discrepancy in the Rx/Tx Byte counters when using the
> i40e driver as a virtual function. Specifically I  am observing that
> the Rx bytes are 4 bytes per packet larger than the Tx Bytes.
> 
> My application is transmitting a packet and then receiving the same
> packet back (in a service function chaining application). I verified
> that the packets being transmitted and received were the same by
> capturing on the interfaces using pdump and analyzing the output in
> Wireshark.
> 
> I believe the discrepancy comes from correcting for the Ethernet FCS.
> I see that in i40e_ethdev.c the CRC bytes are corrected for in the
> i40e_update_vsi_stats() and i40e_read_stats_registers() functions by
> using the following correction (see below). Note the left-hand side
> changes slightly depending on the specific function.
> 
>  pf->internal_stats.rx_bytes -= (pf->internal_stats.rx_unicast +
> pf->internal_stats.rx_multicast + pf->internal_stats.rx_broadcast) *
> ETHER_CRC_LEN;
> 
> However I did not see this correction being used in the vf  driver
> (i40e_ethdev_vf.c). By adding a correction to i40evf_update_stats() as
> follows I can reconcile the counters (and make them consistent with
> the PF behavior).
> 
> nes->rx_bytes -= (nes->rx_unicast + nes->rx_multicast +
> nes->rx_broadcast) * ETHER_CRC_LEN;
> 
> I would like to know if this is the expected behavior or if this is a
> bug in the driver? If this is the expected behavior can you please
> explain why I should expect a discrepancy between the Rx and Tx Byte
> counters?
> 
> 
> Some additional information:
> 
> The application I am observing this behaviour in is using DPDK 17.08.
> (However I don't see any fixes for this in version 17.11).
> 
> The NIC is an Intel Corporation XL710/X710 Virtual Function (rev 02)
> (output from lspci).
> 
> 
> Thank you for your time, and please let me know if you require further
> information.

It is an Intel driver bug.
All drivers must NOT include CRC bytes in counters.
The DPDK (and Linux, BSD) convention is that byte counters do not include
the CRC bytes.  Note: Cisco routers include the CRC bytes; but everyone
else does not.

Re: [dpdk-users] i40e VF Rx/Tx CRC Bytes Discrepancy

[Matthew Coppola]

Thank you for your prompt reply.

BCC'd Users and added dev as I believe it is the more relevant
channel. I apologize for using the wrong channel initially.

I am happy to submit a patch if we would like to resolve the issue. I
am confused why the hardware would have a different behavior for Rx/Tx
Counters. The Tx counters do not contain the CRC bytes but the Rx
counters erroneously(as you pointed out) do. I am having trouble
understanding why the driver makes this discrepancy?

Thank you for your time,

Matt

On Thu, Feb 8, 2018 at 12:32 PM, Stephen Hemminger
<stephen@networkplumber.org> wrote:
> On Thu, 8 Feb 2018 11:52:55 -0500
> Matthew Coppola <matthew.coppola4@gmail.com> wrote:
>
>> Hi,
>>
>> I am noticing a discrepancy in the Rx/Tx Byte counters when using the
>> i40e driver as a virtual function. Specifically I  am observing that
>> the Rx bytes are 4 bytes per packet larger than the Tx Bytes.
>>
>> My application is transmitting a packet and then receiving the same
>> packet back (in a service function chaining application). I verified
>> that the packets being transmitted and received were the same by
>> capturing on the interfaces using pdump and analyzing the output in
>> Wireshark.
>>
>> I believe the discrepancy comes from correcting for the Ethernet FCS.
>> I see that in i40e_ethdev.c the CRC bytes are corrected for in the
>> i40e_update_vsi_stats() and i40e_read_stats_registers() functions by
>> using the following correction (see below). Note the left-hand side
>> changes slightly depending on the specific function.
>>
>>  pf->internal_stats.rx_bytes -= (pf->internal_stats.rx_unicast +
>> pf->internal_stats.rx_multicast + pf->internal_stats.rx_broadcast) *
>> ETHER_CRC_LEN;
>>
>> However I did not see this correction being used in the vf  driver
>> (i40e_ethdev_vf.c). By adding a correction to i40evf_update_stats() as
>> follows I can reconcile the counters (and make them consistent with
>> the PF behavior).
>>
>> nes->rx_bytes -= (nes->rx_unicast + nes->rx_multicast +
>> nes->rx_broadcast) * ETHER_CRC_LEN;
>>
>> I would like to know if this is the expected behavior or if this is a
>> bug in the driver? If this is the expected behavior can you please
>> explain why I should expect a discrepancy between the Rx and Tx Byte
>> counters?
>>
>>
>> Some additional information:
>>
>> The application I am observing this behaviour in is using DPDK 17.08.
>> (However I don't see any fixes for this in version 17.11).
>>
>> The NIC is an Intel Corporation XL710/X710 Virtual Function (rev 02)
>> (output from lspci).
>>
>>
>> Thank you for your time, and please let me know if you require further
>> information.
>
> It is an Intel driver bug.
> All drivers must NOT include CRC bytes in counters.
> The DPDK (and Linux, BSD) convention is that byte counters do not include
> the CRC bytes.  Note: Cisco routers include the CRC bytes; but everyone
> else does not.
>
>