Messages in this thread | | | From | Hariprasad Kelam <> | Subject | Re: [net-next] net: sched: extend flow action with RSS | Date | Mon, 23 Oct 2023 04:04:06 +0000 |
| |
> -----Original Message----- > From: Dave Taht <dave.taht@gmail.com> > ---------------------------------------------------------------------- > On Sat, Oct 21, 2023 at 10:53 PM Sunil Kovvuri <sunil.kovvuri@gmail.com> > wrote: > > > > On Sat, Oct 21, 2023 at 1:43 AM Nambiar, Amritha > > <amritha.nambiar@intel.com> wrote: > > > > > > >> On Fri, Oct 20, 2023 at 2:12 AM Hariprasad Kelam > > > >> <hkelam@marvell.com> > > > >> wrote: > > > >>> > > > >>> This patch extends current flow action with RSS, such that the > > > >>> user can install flower offloads with action RSS followed by a group > id. > > > >>> Since this is done in hardware skip_sw flag is enforced. > > > >> > > > >> Our typical rule for TC is we need s/w equivalence for offloads. > > > >> How would this work in absence of offload? > > > >> > > > > [Hari] > > > > Our typical rule for TC is we need s/w equivalence for offloads. How > would this work in absence of offload? > > > > > > > > This patch we added as an extension to receive queue selection in > hardware. > > > > This patch "act_skbedit: skbedit queue mapping for receive queue" > > > > enabled receive queue selection in hardware and skip_sw is enforced. > > > > > > > > Adding stakeholders of this patch, to get their opinion. > > > > sridhar.samudrala@intel.com amritha.nambiar@intel.com > > > > > > > > incase of RSS, hardware makes decisions about incoming packets > before they are even received in the queue. > > Is there any way to do a LPM to queue match on inbound? > AFAIK, LPM (longest prefix match) is used to make routing decisions by comparing the destination ip address bit-by-bit with prefixes in the routing table. Where on a typice UDP/TCP packet, RSS considers Source IP/Destination IP and Sport and Dport these fields to make decision on the queue mapping.
> > > > > > > > > > The skip_sw for skbedit receive queue action was enforced as the > > > only other alternative was a new hw-only action, or changing the > > > action mirred. See discussion at > > > https://urldefense.proofpoint.com/v2/url?u=https-3A__lore.kernel.org > > > _netdev_20220921132929.3f4ca04d- > 40kernel.org_&d=DwIFaQ&c=nKjWec2b6R0 > > > mOyPaz7xtfQ&r=2bd4kP44ECYFgf- > KoNSJWqEipEtpxXnNBKy0vyoJJ8A&m=_GDltrUw > > > pNUkk5hvYOtlzFeW-rAeZy4_1bSWUdGVen- > sJDjMWkmLw7pVfDH7OHBX&s=SGwkF4_m8 > > > 5hkScL8rNnbVPvDwEnNalysi6x6ual5NHY&e= > > > > > > Few questions WRT this patch: > > > How are the rss groups created? ethtool rss contexts? Any reason to > > > use TC to direct to rss contexts over using ethtool context ids? > > > > > > > Yes, RSS groups are created using ethtool. > > Ethtool ntuple is very limited in expressing flow rules and since the > > general direction is to use 'TC', we are attempting to add RSS action > > to 'TC'. > > > > > > > IIUC, skbedit is meant to only edit skb metadata such as mark, > > > packet type, queue mapping, priority etc. Even if this is a HW only > > > action and has no use in the stack, would skbedit be the right fit here? > > > > > > > The thought was to keep related actions like RQ, RSS group etc under > > one action ie skbedit. > > If that's not the right place we can add a separate action. > > > > Thanks, > > Sunil. > > > > > -- > Oct 30: https://urldefense.proofpoint.com/v2/url?u=https- > 3A__netdevconf.info_0x17_news_the-2Dmaestro-2Dand-2Dthe-2Dmusic- > 2Dbof.html&d=DwIFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=2bd4kP44ECYFgf- > KoNSJWqEipEtpxXnNBKy0vyoJJ8A&m=_GDltrUwpNUkk5hvYOtlzFeW- > rAeZy4_1bSWUdGVen- > sJDjMWkmLw7pVfDH7OHBX&s=gKGp2Lqrx7O6ZAfoWGSQeb__T6PYk8SzKHSf > SxtoNLU&e= > Dave Täht CSO, LibreQos
| |