Re: [dpdk-dev] [PATCH] app/testpmd: fix doubling of 'total TX dropped'

[Andrew Rybchenko <arybchenko@solarflare.com>]

On 7/15/19 5:53 PM, Ferruh Yigit wrote:
> On 7/12/2019 9:32 AM, A.McLoughlin wrote:
>> The 'Accumulated forward statistics for all ports' incorrectly displayed
>> double the actual value for 'total_tx_dropped'. This was because 2
>> lines in the same function both incremented total_tx_dropped every time
>> a packet was dropped.  I removed one of these lines to fix this issue.
>>
>> Fixes: 53324971a14e ("app/testpmd: display/clear forwarding stats on demand")
>> Cc: david.marchand@redhat.com
>> Cc: stable@dpdk.org
>>
>> Signed-off-by: A.McLoughlin <aideen.mcloughlin@intel.com>
>> ---
>>   app/test-pmd/testpmd.c | 1 -
>>   1 file changed, 1 deletion(-)
>>
>> diff --git a/app/test-pmd/testpmd.c b/app/test-pmd/testpmd.c
>> index 3ed3523b7..c41bada50 100644
>> --- a/app/test-pmd/testpmd.c
>> +++ b/app/test-pmd/testpmd.c
>> @@ -1555,7 +1555,6 @@ fwd_stats_display(void)
>>   		total_recv += stats.ipackets;
>>   		total_xmit += stats.opackets;
>>   		total_rx_dropped += stats.imissed;
>> -		total_tx_dropped += ports_stats[pt_id].tx_dropped;
>>   		total_tx_dropped += stats.oerrors;
>>   		total_rx_nombuf  += stats.rx_nombuf;
>>   
>>
> 
> Hi Aideen,
> 
> Indeed 'rte_eth_stats->oerrors' and 'tx_dropped' are different values,
> 
> in testpmd, 'TX-total' is taken as "total_xmit + total_tx_dropped", from this
> description it may be fair to say
> "total_tx_dropped = oerrors + tx_dropped"
> 
> This is easier to see in HW devices, 'oerrors' is the packets sent to HW but HW
> reported failure for some reason, so these packets not transmitted to the medium.
> 'tx_dropped' is mostly calculated by application, driver returns packets that
> can't able to sent to HW, so application can re-try to send or free them and
> increase 'tx_dropped' counter.
> 
> 
> The problem is in the virtual drivers, the packets not able to sent are
> calculated as 'oerrors' and tx_burst functions returns the number of the
> successfully sent packets which cause application calculate remaining ones as
> 'tx_dropped' which cause the duplication.

I don't understand how it is. Tx burst returns a number of owned packets
(either successfully transmitted or internally dropped/freed). If it is 
smaller than number of packets in request, other packets are either 
retried or calculated as tx_dropped.

> In virtual drivers we need to give a decision, if free the wrong packets and
> increase the 'oerrors' or return packets backs to application without increasing
> the 'oerrors', so that application can try to do something with packets or free
> them increasing 'tx_dropped'.
> 
> @David, @Andrew, do you have a suggestion?