[dpdk-dev] [PATCH] latencystats: fix timestamp marking and latency calculation

[longtb5 at viettel.com.vn]

Hi Reshma,

> -----Original Message-----
> From: reshma.pattan at intel.com [mailto:reshma.pattan at intel.com]
> Sent: Friday, September 21, 2018 11:02 PM
> To: longtb5 at viettel.com.vn; konstantin.ananyev at intel.com; dev at dpdk.org
> Cc: Reshma Pattan <reshma.pattan at intel.com>
> Subject: [PATCH] latencystats: fix timestamp marking and latency
calculation
> 
> Latency calculation logic is not correct for the case where packets gets
> dropped before TX. As for the dropped packets, the timestamp is not
> cleared, and such packets still gets counted for latency calculation in
next
> runs, that will result in inaccurate latency measurement.
> 
> So fix this issue as below,
> 
> Before setting timestamp in mbuf, check mbuf don't have any prior valid
> time stamp flag set and after marking the timestamp, set mbuf flags to
> indicate timestamp is valid.
> 
> Before calculating timestamp check mbuf flags are set to indicate
timestamp
> is valid.
> 

This solution as suggested by Konstantin is great. Not only does it solve
the problem but also now the usage of mbuf->timestamp is not exclusive to
latencystats anymore. The application can make use of timestamp at the same
as latencystats simply by toggling PKT_RX_TIMESTAMP. I think we should
update the doc to include this information. 

> With the above logic it is guaranteed that correct timestamps have been
> used.
> 
> Fixes: 5cd3cac9ed ("latency: added new library for latency stats")
> 
> Reported-by: Bao-Long Tran <longtb5 at viettel.com.vn>
> Signed-off-by: Reshma Pattan <reshma.pattan at intel.com>
> ---
>  lib/librte_latencystats/rte_latencystats.c | 8 ++++++--
>  1 file changed, 6 insertions(+), 2 deletions(-)
> 
> diff --git a/lib/librte_latencystats/rte_latencystats.c
> b/lib/librte_latencystats/rte_latencystats.c
> index 1fdec68e3..8870226bb 100644
> --- a/lib/librte_latencystats/rte_latencystats.c
> +++ b/lib/librte_latencystats/rte_latencystats.c
> @@ -125,8 +125,11 @@ add_time_stamps(uint16_t pid __rte_unused,
>  	for (i = 0; i < nb_pkts; i++) {
>  		diff_tsc = now - prev_tsc;
>  		timer_tsc += diff_tsc;
> -		if (timer_tsc >= samp_intvl) {
> +
> +		if ((pkts[i]->ol_flags & PKT_RX_TIMESTAMP) == 0
> +			&& (timer_tsc >= samp_intvl)) {
>  			pkts[i]->timestamp = now;
> +			pkts[i]->ol_flags |= PKT_RX_TIMESTAMP;
>  			timer_tsc = 0;
>  		}
>  		prev_tsc = now;
> @@ -156,7 +159,8 @@ calc_latency(uint16_t pid __rte_unused,
> 
>  	now = rte_rdtsc();
>  	for (i = 0; i < nb_pkts; i++) {
> -		if (pkts[i]->timestamp)
> +		if ((pkts[i]->ol_flags & PKT_RX_TIMESTAMP) &&
> +				pkts[i]->timestamp)

Just a nit, but I think we don't have to check for pkts[i]->timestamp here.

>  			latency[cnt++] = now - pkts[i]->timestamp;
>  	}
> 
> --
> 2.14.4

Best regards,
BL