[dpdk-stable] [dpdk-dev] [PATCH v2] net/af_packet: improve Tx statistics accuracy

[Musatescu, Flavia]

On 10/10/2019 19:35, Stephen Hemminger wrote:
> On Thu, 10 Oct 2019 18:34:30 +0100
> Flavia Musatescu <flavia.musatescu at intel.com> wrote:
>
>> When sendto call fails and ENOBUFS error is being set some of the
>> packets are actually successfully transmitted. There is no available
>> count of those packets, so in order to make the statistics more
>> accurate, all the previously enqueued packets will be considered
>> successful, even though this is not entirely correct.
>>
>> Before:
>>     testpmd Tx statistics:
>>        TX-packets: 7529062   TX-errors: 3702150  TX-bytes:  451743720
>>     pktgen Rx statistics:
>>        Total Rx Pkts: 10700700
>>
>> After:
>>     testpmd TX statistics:
>>        TX-packets: 11510625  TX-errors: 0        TX-bytes:  690637500
>>     pktgen Rx statistics:
>>        Total Rx Pkts: 10974307
>>
>> Fixes: 74b7fc0a0ff1 ("net/af_packet: fix packet bytes counting")
>> Cc: ciwillia at brocade.com
>> Cc: stable at dpdk.org
>>
>> Signed-off-by: Flavia Musatescu <flavia.musatescu at intel.com>
>>
>> ---
>> v2:
>> * Changed the comment
>> ---
>>   drivers/net/af_packet/rte_eth_af_packet.c | 8 ++++++--
>>   1 file changed, 6 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/net/af_packet/rte_eth_af_packet.c b/drivers/net/af_packet/rte_eth_af_packet.c
>> index 6df09f2..df281bf 100644
>> --- a/drivers/net/af_packet/rte_eth_af_packet.c
>> +++ b/drivers/net/af_packet/rte_eth_af_packet.c
>> @@ -244,8 +244,12 @@ eth_af_packet_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
>>   	}
>>   
>>   	/* kick-off transmits */
>> -	if (sendto(pkt_q->sockfd, NULL, 0, MSG_DONTWAIT, NULL, 0) == -1) {
>> -		/* error sending -- no packets transmitted */
>> +	if (sendto(pkt_q->sockfd, NULL, 0, MSG_DONTWAIT, NULL, 0) == -1 &&
>> +			errno != ENOBUFS) {
>> +		/*
>> +		 * In case of a ENOBUFS error all of the enqueued packets will
>> +		 * be considered successful even though only some are sent.
>> +		 */
>>   		num_tx = 0;
>>   		num_tx_bytes = 0;
>>   	}
> What about EINTR or EAGAIN?
Hi Stephen,

I agree with EAGAIN error case as I was able to test this scenario and 
check the statistics.

Can we get an EINTR error in case of a nonblocking operation 
(MSG_DONTWAIT flag is being used)? Is this a common situation and will 
the gain justify the cost of an additional check? Do you have any 
suggestions for how I could test this scenario?

Thanks,
Flavia