Build and debug DOCA Flow packet-steering applications on NVIDIA NICs and DPUs.
npx skills add https://github.com/NVIDIA/skills --skill doca-flow
Please help me install this Agent Skill. Skill: doca-flow Source: https://github.com/NVIDIA/skills/tree/main/skills/doca-flow Command: npx skills add https://github.com/NVIDIA/skills --skill doca-flow
Programming hardware packet steering with the DOCA Flow API is error-prone: agents invent nonexistent API names, misconfigure port bring-up, and silently fall back to kernel tc/iptables instead of libdoca_flow. This Skill grounds every step in the installed DOCA headers and shipped samples so pipes, ports, and counters actually program the hardware.
Ask your agent to bring up a DOCA Flow port on your NIC and create a pipe that matches IPv4 traffic and forwards it to a representor with per-entry counters.
High-intent search queries and answers about installing and using this skill.
First check device placement with mlxconfig: in SEPARATED_HOST mode the steering plane belongs to the DPU Arm cores, so host-side port start fails. Then allocate hugepages, probe devices with dv_flow_en=2, and call doca_flow_port_start with a mandatory port_id and device source.
Build a doca_flow_pipe_cfg with explicit match and action masks, then call doca_flow_pipe_create, which validates the spec at construction time. Add entries with the pipe-type-specific call such as doca_flow_pipe_basic_add_entry, copying the argument list from the installed header.
This signature means the opened function does not own the hardware steering plane, typically because a BlueField is in SEPARATED_HOST (NIC) mode where steering belongs to the DPU Arm. Run Flow on the Arm side or change the card mode via mlxconfig; no pipe-spec edit fixes this.
Yes, the flow-ct module (doca_flow_ct.h) ships inside doca-flow and provides hardware-accelerated 5-tuple connection tracking, aging, and SNAT/DNAT. Initialize it after doca_flow_init but before port start, and confirm device support with doca_flow_ct_cap_is_dev_supported.
Not within this Skill's scope: kernel tc/flower, iptables, eBPF/XDP, OVS, and bare rte_flow bypass the DOCA Flow model of pipes, entries, and hardware counters. The Skill routes those questions elsewhere and insists deliverables link libdoca_flow.
Counters require two steps: reserving a global pool via doca_flow_cfg_set_nr_counters at init, and passing a populated monitor struct to the add-entry call. Passing NULL for the monitor adds the entry successfully but binds no counter, so queries return zero under traffic.
yangwhaleInstall, configure, and troubleshoot DeepEP on NVIDIA GPU systems.
psaabValidate GRE userspace dataplane stability and redundancy-group failover with iperf3.
stolyarchukGuide C++ engineers in designing and optimizing DOCA Flow pipelines on NVIDIA SmartNICs.
openeuler-mirrorDiagnose Linux Netfilter, iptables, nftables, and conntrack connectivity issues by correlating rule chains with connection states.
