On 10 Nov 2026, at 17:08, Eli Britstein wrote:
> On 09/11/2026 17:43, Eelco Chaudron wrote: >> External email: Use caution opening links or attachments >> >> >> Allow the netdev_hw_post_process() callback to return an optional set of >> alternative actions via a new 'alt_actions' parameter. When set, these >> actions are executed instead of the flow's own actions, enabling offload >> providers to indicate that a subset of the actions was already performed >> in hardware. >> >> v4 changes: >> - Add additional counter to count action continuation hits. > > Hi Eelco, thanks for the counter enhancement. > > However, my meaning about statistics was for per flow. > > This feature is aimed to be used in CT offload. With such, consider a > typical flow: > > <M>,actions=A1,ct(),recirc() > > If all is offloaded (both the megaflow and the connection), we will hit fully > in HW. > > With a new connection, the same megaflow is in HW. The HW will match <M> and > perform A1, but we will miss in CT, so we will continue > > the processing in SW. > > It will be useful if we have such info in ovs-appctl dpctl/dump-flows -m. > > For example (from our downstream): > > ufid:fe4f3bd1-7a31-42ea-9a0a-a54dfd363f3f, > recirc_id(0),dp_hash(0/0),skb_priority(0/0),nv_mp_pid(0/0),nv_mp_preferred(0/0),nv_mp_strict(0/0),in_port(enp8s0f0_0),skb_mark(0/0),ct_state(0/0x20),ct_zone(0/0),ct_mark(0/0),ct_label(0/0),packet_type(ns=0,id=0),eth(src=00:00:00:00:00:01/00:00:00:00:00:00,dst=00:00:00:00:00:02/00:00:00:00:00:00),eth_type(0x0800),ipv4(src=1.1.1.1/0.0.0.0,dst=2.2.2.2/0.0.0.0,proto=17/0,tos=0/0,ttl=64/0,frag=no),udp(src=1000/0,dst=2000/0), > partial_offload_packets:4, partial_offload_bytes:240, packets:7, bytes:420, > used:2.185s, offloaded:yes, dp:doca, actions:ct(zone=2),recirc(0x1), > dp-extra-info:miniflow_bits(3,2) > > As I mentioned, we removed altogether the support for the "legacy" partial > offload (match only), so our "partial" meaning is only for this case. > > For upstream, I think we should decide on names. > > > Another enhancement I suggested (same area, not those commits) is to have > (per-flow, like this), is a separation between SW/HW counters. > > It can show both (even for a simple flow), in case the offload thread takes > some time while some packets still pass in SW. It's an edge case, but still > worth considering. Hi Eli, I got your point from our discussion, but I thought it would be good for now to have a separate counter on the PMD level. Doing hw vs sw counters is on my wish list for a while. Actually it's already there in the kernel for TC, and we get them. There is currently no way to display them. I've created a todo item for myself. When I get time I'll pick it up, but it's low priority. Cheers, Eelco >> >> v3 changes: >> - Reworked dummy_build_alt_actions() to preserve action ordering, >> processing actions up to the first non-offloadable port and >> returning the remainder as alt_actions. Also changed >> dummy_offload_hw_process_pkt() to adhere to this. >> - Restored comment in dummy_netdev_simulate_offload() that was >> incorrectly removed. >> - Fixed potential ABBA deadlock in dummy_netdev_hw_offload_run() >> that was introduced. >> >> v2 changes: >> - Removed first patch adding the ALWAYS_INLINE attribute >> - Use positive error codes in dpif-offload-dummy.c >> - Fixed some style issues >> >> Eelco Chaudron (2): >> dpif-offload: Add support for alternative actions. >> dpif-offload-dummy: Exercise hw_post_process() alt_actions API. >> >> lib/dpif-netdev-perf.c | 3 + >> lib/dpif-netdev-perf.h | 2 + >> lib/dpif-netdev.c | 113 ++++++++++++++++++++------- >> lib/dpif-offload-dpdk.c | 5 +- >> lib/dpif-offload-dummy.c | 148 +++++++++++++++++++++++++++++++----- >> lib/dpif-offload-provider.h | 7 +- >> lib/dpif-offload.c | 4 +- >> lib/dpif-offload.h | 7 ++ >> tests/dpif-netdev.at | 103 ++++++++++++++++++++++++- >> tests/ofproto-dpif.at | 65 ++++++++++------ >> tests/pmd.at | 3 +- >> 11 files changed, 384 insertions(+), 76 deletions(-) >> >> -- >> 2.56.0 _______________________________________________ dev mailing list [email protected] https://mail.openvswitch.org/mailman/listinfo/ovs-dev
