[dpdk-dev] [PATCH v3] examples/ipsec-secgw: add per core packet stats

[Ananyev, Konstantin]

> > > @@ -1099,6 +1151,10 @@ ipsec_poll_mode_worker(void)
> > >  	const uint64_t drain_tsc = (rte_get_tsc_hz() + US_PER_S - 1)
> > >  			/ US_PER_S * BURST_TX_DRAIN_US;
> > >  	struct lcore_rx_queue *rxql;
> > > +#if (STATS_INTERVAL > 0)
> > > +	const uint64_t timer_period = STATS_INTERVAL * rte_get_timer_hz();
> > > +	uint64_t timer_tsc = 0;
> > > +#endif /* STATS_INTERVAL */
> > >
> > >  	prev_tsc = 0;
> > >  	lcore_id = rte_lcore_id();
> > > @@ -1159,6 +1215,19 @@ ipsec_poll_mode_worker(void)
> > >  			drain_tx_buffers(qconf);
> > >  			drain_crypto_buffers(qconf);
> > >  			prev_tsc = cur_tsc;
> > > +#if (STATS_INTERVAL > 0)
> > > +			if (lcore_id == rte_get_master_lcore()) {
> > > +				/* advance the timer */
> > > +				timer_tsc += diff_tsc;
> > > +
> > > +				/* if timer has reached its timeout */
> > > +				if (unlikely(timer_tsc >= timer_period)) {
> > > +					print_stats();
> > > +					/* reset the timer */
> > > +					timer_tsc = 0;
> > > +				}
> > > +			}
> > > +#endif /* STATS_INTERVAL */
> >
> > I still don't understand why to do it in data-path thread.
> > As I said in previous comments, in DPDK there is a control thread that can be
> > used for such house-keeping tasks.
> > Why not to use it (via rte_alarm or so) and keep data-path threads less affected.
> 
> [Anoob] From Marvell's estimates, this stats collection and reporting will be expensive and so cannot be enabled by default. This is required
> for analyzing the traffic distribution in cases where the performance isn't scaling as expected.

Understood.

 > And this patch achieves the desired feature. 

Ok, but why not to do it in control (house-keeping) thread?
That would achieve desired goal and keep data-path unaffected.

> If Intel would like to improve the approach, that can be taken up as a separate patch.

This is not a vendor specific part. 
You making changes in common data-path code that is used by all ipsec-secgw users.
I think it is everyone benefit (and responsibility) to keep common data-path
code clean, tidy and fast.
If we can avoid polluting it with extra code, I don't see a reason not to do it. 

> 
> >
> > >  		}
> > >
> > >  		for (i = 0; i < qconf->nb_rx_queue; ++i) { @@ -1169,8 +1238,10
> > @@
> > > ipsec_poll_mode_worker(void)
> > >  			nb_rx = rte_eth_rx_burst(portid, queueid,
> > >  					pkts, MAX_PKT_BURST);
> > >
> > > -			if (nb_rx > 0)
> > > +			if (nb_rx > 0) {
> > > +				core_stats_update_rx(nb_rx);
> > >  				process_pkts(qconf, pkts, nb_rx, portid);
> > > +			}
> > >
> > >  			/* dequeue and process completed crypto-ops */
> > >  			if (is_unprotected_port(portid))
> > > diff --git a/examples/ipsec-secgw/ipsec-secgw.h
> > > b/examples/ipsec-secgw/ipsec-secgw.h
> > > index 4b53cb5..5b3561f 100644
> > > --- a/examples/ipsec-secgw/ipsec-secgw.h
> > > +++ b/examples/ipsec-secgw/ipsec-secgw.h
> > > @@ -6,6 +6,8 @@
> > >
> > >  #include <stdbool.h>
> > >
> > > +#define STATS_INTERVAL 0
> >
> > Shouldn't it be:
> > #ifndef STATS_INTERVAL
> > #define STATS_INTERVAL	0
> > #endif
> > ?
> 
> [Anoob] Will update in v4.
> 
> >
> > To allow user specify statis interval via EXTRA_CFLAGS='-DSTATS_INTERVAL=10'
> > or so.
> >
> > > +
> > >  #define NB_SOCKETS 4
> > >
> > >  #define MAX_PKT_BURST 32
> > > @@ -69,6 +71,17 @@ struct ethaddr_info {
> > >  	uint64_t src, dst;
> > >  };
> > >
> > > +#if (STATS_INTERVAL > 0)
> > > +struct ipsec_core_statistics {
> > > +	uint64_t tx;
> > > +	uint64_t rx;
> > > +	uint64_t dropped;
> > > +	uint64_t burst_rx;
> > > +} __rte_cache_aligned;
> > > +
> > > +struct ipsec_core_statistics core_statistics[RTE_MAX_LCORE]; #endif
> > > +/* STATS_INTERVAL */
> > > +
> > >  extern struct ethaddr_info ethaddr_tbl[RTE_MAX_ETHPORTS];
> > >
> > >  /* Port mask to identify the unprotected ports */ @@ -85,4 +98,59 @@
> > > is_unprotected_port(uint16_t port_id)
> > >  	return unprotected_port_mask & (1 << port_id);  }
> > >
> > > +static inline void
> > > +core_stats_update_rx(int n)
> > > +{
> > > +#if (STATS_INTERVAL > 0)
> > > +	int lcore_id = rte_lcore_id();
> > > +	core_statistics[lcore_id].rx += n;
> > > +	if (n == MAX_PKT_BURST)
> > > +		core_statistics[lcore_id].burst_rx += n; #else
> > > +	RTE_SET_USED(n);
> > > +#endif /* STATS_INTERVAL */
> > > +}
> > > +
> > > +static inline void
> > > +core_stats_update_tx(int n)
> > > +{
> > > +#if (STATS_INTERVAL > 0)
> > > +	int lcore_id = rte_lcore_id();
> > > +	core_statistics[lcore_id].tx += n;
> > > +#else
> > > +	RTE_SET_USED(n);
> > > +#endif /* STATS_INTERVAL */
> > > +}
> > > +
> > > +static inline void
> > > +core_stats_update_drop(int n)
> > > +{
> > > +#if (STATS_INTERVAL > 0)
> > > +	int lcore_id = rte_lcore_id();
> > > +	core_statistics[lcore_id].dropped += n; #else
> > > +	RTE_SET_USED(n);
> > > +#endif /* STATS_INTERVAL */
> > > +}
> > > +
> > > +/* helper routine to free bulk of packets */ static inline void
> > > +free_pkts(struct rte_mbuf *mb[], uint32_t n) {
> > > +	uint32_t i;
> > > +
> > > +	for (i = 0; i != n; i++)
> > > +		rte_pktmbuf_free(mb[i]);
> > > +
> > > +	core_stats_update_drop(n);
> > > +}
> > > +
> > > +/* helper routine to free single packet */ static inline void
> > > +free_pkt(struct rte_mbuf *mb) {
> > > +	rte_pktmbuf_free(mb);
> > > +	core_stats_update_drop(1);
> >
> > Probably just:
> > free_pkts(&mb, 1);
> > ?
> 
> [Anoob] Will update in v4.
> 
> >
> > > +}
> > > +
> > >  #endif /* _IPSEC_SECGW_H_ */