RE: [EXT] Re: [PATCH] examples/ipsec-secgw: fix IPsec performance drop

[Rahul Bhansali <rbhansali@marvell.com>]

> -----Original Message-----
> From: Ferruh Yigit <ferruh.yigit@amd.com>
> Sent: Tuesday, February 6, 2024 11:55 PM
> To: Rahul Bhansali <rbhansali@marvell.com>; dev@dpdk.org; Radu Nicolau
> <radu.nicolau@intel.com>; Akhil Goyal <gakhil@marvell.com>; Konstantin
> Ananyev <konstantin.ananyev@huawei.com>; Anoob Joseph
> <anoobj@marvell.com>
> Subject: [EXT] Re: [PATCH] examples/ipsec-secgw: fix IPsec performance drop
> 
> External Email
> 
> ----------------------------------------------------------------------
> On 2/6/2024 12:38 PM, Rahul Bhansali wrote:
> > Single packet free using rte_pktmbuf_free_bulk() is dropping the
> > performance. On cn10k, maximum of ~4% drop observed for IPsec event
> > mode single SA outbound case.
> >
> > To fix this issue, single packet free will use rte_pktmbuf_free API.
> >
> > Fixes: bd7c063561b3 ("examples/ipsec-secgw: use bulk free")
> >
> > Signed-off-by: Rahul Bhansali <rbhansali@marvell.com>
> > ---
> >  examples/ipsec-secgw/ipsec-secgw.h | 7 +++----
> >  1 file changed, 3 insertions(+), 4 deletions(-)
> >
> > diff --git a/examples/ipsec-secgw/ipsec-secgw.h
> > b/examples/ipsec-secgw/ipsec-secgw.h
> > index 8baab44ee7..ec33a982df 100644
> > --- a/examples/ipsec-secgw/ipsec-secgw.h
> > +++ b/examples/ipsec-secgw/ipsec-secgw.h
> > @@ -229,11 +229,10 @@ free_reassembly_fail_pkt(struct rte_mbuf *mb)  }
> >
> >  /* helper routine to free bulk of packets */ -static inline void
> > -free_pkts(struct rte_mbuf *mb[], uint32_t n)
> > +static __rte_always_inline void
> > +free_pkts(struct rte_mbuf *mb[], const uint32_t n)
> >  {
> > -	rte_pktmbuf_free_bulk(mb, n);
> > -
> > +	n == 1 ? rte_pktmbuf_free(mb[0]) : rte_pktmbuf_free_bulk(mb, n);
> >  	core_stats_update_drop(n);
> >  }
> >
> 
> Hi Rahul,
> 
> Do you think the 'rte_pktmbuf_free_bulk()' API performance can be improved by
> similar change?

Hi Ferruh,
Currently 'rte_pktmbuf_free_bulk() is not inline. If we make that along with __rte_pktmbuf_free_seg_via_array()  both inline then performance can be improved similar.