Re: [dpdk-dev] [dpdk-stable] [PATCH v3 14/22] net/ena: indicate Rx RSS hash presence

["Michał Krawczyk" <mk@semihalf.com>]

pt., 7 maj 2021 o 18:47 Ferruh Yigit <ferruh.yigit@intel.com> napisał(a):
>
> On 5/6/2021 3:25 PM, Michal Krawczyk wrote:
> > To make it possible to the app to determine if the hash was calculated
> > for the packet or not, the PKT_RX_RSS_HASH should be set in the mbuf's
> > ol_flags.
> >
> > As the PMD wasn't setting that, the application couldn't check if there
> > is a hash in a proper way.
> >
> > The hash is valid only if it's UDP or TCP and the IP packet wasn't
> > fragmented.
> >
> > Fixes: e5df9f33db00 ("net/ena: fix passing RSS hash to mbuf")
> > Cc: stable@dpdk.org
> >
> > Signed-off-by: Michal Krawczyk <mk@semihalf.com>
> > Reviewed-by: Igor Chauskin <igorch@amazon.com>
> > Reviewed-by: Amit Bernstein <amitbern@amazon.com>
> > ---
> >  drivers/net/ena/ena_ethdev.c | 9 +++++++--
> >  1 file changed, 7 insertions(+), 2 deletions(-)
> >
> > diff --git a/drivers/net/ena/ena_ethdev.c b/drivers/net/ena/ena_ethdev.c
> > index 6092288239..5d107775f4 100644
> > --- a/drivers/net/ena/ena_ethdev.c
> > +++ b/drivers/net/ena/ena_ethdev.c
> > @@ -51,6 +51,8 @@
> >
> >  #define ENA_MIN_RING_DESC    128
> >
> > +#define ENA_PTYPE_HAS_HASH   (RTE_PTYPE_L4_TCP | RTE_PTYPE_L4_UDP)
> > +
> >  enum ethtool_stringset {
> >       ETH_SS_TEST             = 0,
> >       ETH_SS_STATS,
> > @@ -314,6 +316,11 @@ static inline void ena_rx_mbuf_prepare(struct rte_mbuf *mbuf,
> >               else
> >                       ol_flags |= PKT_RX_L4_CKSUM_GOOD;
> >
> > +     if (likely((packet_type & ENA_PTYPE_HAS_HASH) && !ena_rx_ctx->frag)) {
> > +             ol_flags |= PKT_RX_RSS_HASH;
> > +             mbuf->hash.rss = ena_rx_ctx->hash;
> > +     }
>
> The driver should announce the 'DEV_RX_OFFLOAD_RSS_HASH' offload capability
> first, and should set 'PKT_RX_RSS_HASH' only if application requests the
> 'DEV_RX_OFFLOAD_RSS_HASH'.
>
>
> Normally this is an optimization to not always update the 'mbuf->hash.rss' but
> wait for explicit request from application for it.
>
> But there is no practical performance gain for some PMDs and they enable it even
> user does not ask for it [1] (of course if RSS is enabled), if this is same for
> 'ena' you can do the same.
>

Thanks for the reference - ENA should do the same, I will change that in v4.

> [1]
> https://git.dpdk.org/dpdk/tree/drivers/net/ixgbe/ixgbe_ethdev.c?h=v21.05-rc2#n2384
>
> > +
> >       mbuf->ol_flags = ol_flags;
> >       mbuf->packet_type = packet_type;
> >  }
> > @@ -2245,8 +2252,6 @@ static uint16_t eth_ena_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
> >                       ++rx_ring->rx_stats.bad_csum;
> >               }
> >
> > -             mbuf->hash.rss = ena_rx_ctx.hash;
> > -
> >               rx_pkts[completed] = mbuf;
> >               rx_ring->rx_stats.bytes += mbuf->pkt_len;
> >       }
> >
>