All skills
nvidia avatar

/doca-programming-guide

@f64fa0e
by NVIDIA Corporationnvidia/skills3.5k stars
424

Use this skill when the user is writing their first DOCA app or asking a library-agnostic programming question — picking a shipped sample to copy and modify, wiring the canonical pkg-config doca-{library} + meson build (or FFI from Rust / Go / Python against the public C ABI), walking the cfg-create → init → start → use → stop → destroy lifecycle, validating a spec before commit, or decoding a DOCA_ERROR_* return with doca_error_get_descr(). Trigger even when the user does not say "DOCA programming guide" — implicit phrasings: "write my first DOCA program", "meson line for doca_rdma_*", "got DOCA_ERROR_BAD_STATE on my first call", "call DOCA from Rust without writing C", "built clean but nothing on the wire", "what order do doca_*_pipe calls go in". Refuse and route for install / hugepages / pkg-config not resolving doca-{library} (doca-setup), docs or version lookup (doca-public-knowledge-map), and library-internal API construction like Flow pipe topology or RDMA QP setup (matching library skill).

Use this Skill: https://skilld.dev/gh/nvidia/skills/doca-programming-guide

This session only. Nothing lands on disk.

CAPABILITIES.md

≈4.4k tokens on demand. Your agent reads this file only when SKILL.md points to it.

DOCA programming guide — capabilities, version compatibility, errors, observability, safety

Where to start: Read ## Pattern overview first to pick the family; then drill into the matching H2. Every prescriptive section below stays library-agnostic — overlays for a specific library live in that library's CAPABILITIES.md.

Read this file when the loader sent you here from SKILL.md. For the step-by-step workflows that use the surface described here, see TASKS.md. For env-class equivalents (install profiles, env-class errors, env observability), see doca-setup. For where to find official documentation, the on-disk install layout, or the Installation Guide, route through doca-public-knowledge-map.

This file describes the shape every DOCA program shares — the architecture every library plugs into, the lifecycle every library follows, the error pattern every library returns, the observability surface every library emits into, and the safety rules every program must respect. Library-specific overlays live in the matching library skill.

Lint note: this file references /opt/mellanox/doca and docs.nvidia.com paths several times. The lint warns about that for most library skills and asks for cross-links to doca-public-knowledge-map instead. For doca-programming-guide specifically the references are intrinsic: the skill's whole job is to describe how a DOCA program is shaped on disk and on the wire, and the canonical install tree (/opt/mellanox/doca) plus the canonical docs tree (docs.nvidia.com/doca/sdk/) are the two coordinate systems every prescription is anchored against. The repeated paths are intentional; do not refactor them out.

Pattern overview

Every program-class question this skill teaches reduces to one of FIVE patterns. Reach for the pattern first; the H2 below it carries the substance.

Pattern Class shape Where it lives
1. Plug into the right slot Host / DPU / Switch; library / app / service / tool — orient the program ## Capabilities and modes
2. Follow the universal lifecycle cfg → create → init → start → use → stop → destroy for every DOCA object ## Capabilities and modes lifecycle subsection
3. Build / link the program pkg-config doca-<library> (C/C++ direct) OR FFI/bindings against the public C ABI TASKS.md ## build
4. Handle DOCA_ERROR_* correctly doca_error_get_descr() first; never paper over ## Error taxonomy
5. Validate before commit, observe before "fixing" Pre-commit validation, log levels, sample-counter discipline ## Safety policy + ## Observability

Two principles cut across all five:

  • Library-agnostic shape, library-specific overlay. This skill prescribes the shape; the matching library skill (e.g. doca-flow) overlays the library-specific specifics (Flow pipe topology, RDMA QP semantics, …). Both must be loaded for a complete answer.
  • Programs run with DOCA, not of DOCA. Everything below assumes the user is calling doca_* from their own code. Patches to DOCA itself are out of scope and belong in the DOCA contributor repo, not this bundle.

Capabilities and modes

DOCA is structured into a small set of orthogonal pieces. The agent must clarify which piece the user is touching before recommending any code or command.

Where the code runs. The same DOCA library can be linked from a program running in any of three locations. Which modes are available on the user's host is an env property and lives in doca-setup CAPABILITIES.md ## Capabilities and modes; which mode the program selects at init time is a programming property and is described here.

Location What it is When the user's program is here
Host x86 / Arm host with a BlueField as a SmartNIC, or a host with a ConnectX NIC Most C / C++ first-app development; building, linking, and running on a developer workstation. The program sees physical ports of the BlueField from the host's view, via PF/VF representors.
DPU Inside the BlueField OS itself (Arm). Workloads that have to run on the DPU rather than across PCIe. The program sees local network interfaces and (in switch mode) representors of the host's PF/VF.
Switch DPU-side, with the BlueField in switch (DPU) mode. Multi-tenant steering, full virtual switch on the DPU. The program takes ownership of representors; library-imposed constraints apply (e.g. Flow forbids rte_eth_dev_* configure/start on representors in switch mode — Flow takes them over).

The library guides describe which mode is meaningful for that library; the agent must clarify which of the three the user's program is in before recommending any port-id, representor, or queue naming.

What kind of artifact you're building. DOCA ships four categories of code; the build / run / debug pattern depends on which one the user's project is.

Category Where it lives in an installed tree What the user does with it
Library /opt/mellanox/doca/lib/<arch>-linux-gnu/libdoca_<library>.so The user's program links against this via pkg-config doca-<library>. The first-app workflow (TASKS.md ## modify) almost always derives from a sample that consumes one library.
Sample /opt/mellanox/doca/samples/doca_<library>/<sample_name>/ One library, minimal scaffolding. The canonical modify-a-shipped-sample starting point.
Reference application /opt/mellanox/doca/applications/<app>/ Larger, integrated example combining multiple libraries (e.g. Flow + Telemetry). Useful as a read-only worked example; rarely the right starting point for a first app.
Service / tool /opt/mellanox/doca/services/, /opt/mellanox/doca/tools/ Long-running processes (services) or CLIs (tools like doca_caps). Not the user's program; the user's program consumes their output.

The agent's rule: when the user says "first app", the answer always starts from a sample, not from an application or a service. Applications are for reading; samples are for modifying.

Build flavor selection. The pkg-config module name selects which library variant your program links against — same library, different runtime properties. The env-side mechanics (where the trace *.so lives on disk, how to set LD_LIBRARY_PATH) are in doca-setup CAPABILITIES.md ## Capabilities and modes; the program-side rationale for which flavor to pick is here.

pkg-config module Variant When to pick it
doca-<library> (e.g. doca-flow) Release. Optimized, no extra sanitation. Production and benchmarking.
doca-<library>-trace (e.g. doca-flow-trace) Trace build. Enables additional input sanitation, slower, much louder logs. First-app development. First-time-user errors are caught earlier and surfaced with explicit reasons; switch to release only after the staged run succeeds and you're moving to a performance reading.

The universal lifecycle. Every DOCA library object the user creates follows the same six-phase lifecycle. Library skills name which specific API calls correspond to each phase.

Phase What it means What goes wrong if you skip it
cfg-create Allocate a configuration object (doca_<library>_cfg_*). NULL deref on the cfg pointer.
cfg-set-* Populate the configuration with the user's settings (mode, capabilities, callbacks). The library defaults what you didn't set — usually not what you meant.
init / create Materialize the actual library object from the populated cfg. Calls that need the object return DOCA_ERROR_BAD_STATE.
start Move the object into a state where it can be used. use-phase calls return DOCA_ERROR_BAD_STATE.
use The library-specific work — pipe entries, queue ops, channel sends, etc. Library-specific symptoms (no traffic, dropped packets, hung sends).
stop + destroy Reverse-order tear-down; release in the inverse order of acquisition. Resource leaks; on next run, capability discovery may report exhausted budgets.

The agent's rule: never recommend a use-phase call before confirming the object is in started state, and never recommend re-creating an object that hasn't been destroyed.

Version compatibility

For the canonical DOCA version-detection chain, the four-way match rule, NGC container semantics, the headers-win-over-docs rule, and the routing to the DOCA Compatibility Policy, see doca-version. The body lives there; this skill does not duplicate it.

The program-side overlay layers on top of the verified version:

  • Quote the version observed, never "latest" and never agent memory. API names, sample filenames, and capability availability all depend on the user's installed version. Per the safety policy in doca-version CAPABILITIES.md ## Safety policy, a recommendation that quotes a version-pinned URL or invents a feature based on agent memory is the bundle's primary hallucination failure mode.
  • For non-C consumers (FFI / bindings). The binding tooling generates declarations from the headers at build time. Re-generate after any DOCA upgrade — bindings pinned to a stale header will compile but call into the new *.so with mismatched signatures, returning DOCA_ERROR_INVALID_VALUE or, worse, silent corruption.
  • Use the per-library doca_<library>_cap_* query family at runtime as the authoritative answer for "is this capability supported on this device + this install". The version-matrix is the promise; the cap query is the reality. When they disagree, the cap query wins — see doca-version TASKS.md ## test step 3.

Error taxonomy

Every DOCA library returns a doca_error_t from any call that can fail. The error family is library-agnostic at the type level; library-specific overlays (e.g. which DOCA_ERROR_* Flow returns from which call, with what recovery) live in the matching library skill.

The general taxonomy the agent must recognize:

doca_error_t value Meaning Programming-side response
DOCA_SUCCESS The call succeeded. Continue.
DOCA_ERROR_INVALID_VALUE One of the arguments is malformed (NULL where required, out-of-range, mismatched type). The bug is in the program, not the library. Check the call's documentation in the library guide; correct the argument; re-validate.
DOCA_ERROR_BAD_STATE The object is not in a state where this call is meaningful (e.g. start before init, use before start). Lifecycle order is wrong. Walk the call sequence against the universal lifecycle in ## Capabilities and modes; fix the order.
DOCA_ERROR_NOT_SUPPORTED The hardware / firmware / mode does not implement this feature on this host. Re-run capability discovery (doca_caps, the library's own capability-query API). The agent must not attempt code workarounds; either change the program's intent, switch hardware/mode, or escalate.
DOCA_ERROR_NO_MEMORY Allocation failed. Out-of-memory at the system level (rare during first-app development) or out of a library-managed pool. The library skill names the relevant pool / budget.
DOCA_ERROR_TIME_OUT An async operation did not complete in time. Increase the timeout if intended; investigate why the underlying operation is slow if not.
DOCA_ERROR_INITIALIZATION Library / object initialization itself failed. Almost always a missing prerequisite (driver, capability, peer). Walk through the library's ## configure workflow; do not retry blindly.
DOCA_ERROR_DRIVER The underlying driver / firmware reported failure. Capture state and route to env-class debug (doca-setup ## debug) — the layer below DOCA is the suspect, not the program.
DOCA_ERROR_UNKNOWN Internal error not mappable to the taxonomy. Capture state and escalate through doca-debug ## debug; do not paper over with a retry loop.

For any returned doca_error_t other than DOCA_SUCCESS, the agent's first step is to call doca_error_get_descr(err) and quote the actual descriptor string — never paraphrase from memory. The descriptor is the library's own statement of what went wrong; library-specific guides can refine it but cannot contradict it.

The env-class error surface (missing *.pc, hugepages not mounted, representor not visible) is separate and lives in doca-setup CAPABILITIES.md ## Error taxonomy. The agent's debug order is always env first (doca-setup ## debug layers 1–4), then program (TASKS.md ## debug layers 2–5).

Observability

Every DOCA program has access to the same shared observability surface; library-specific surfaces (Flow counters, RDMA queue stats, …) extend it.

Logging. DOCA libraries log via the doca_log_register_* API and emit to stderr by default. The CLI flag --sdk-log-level <level> (or the equivalent programmatic call doca_log_backend_set_sdk_level()) controls verbosity:

  • 30 (DOCA_LOG_LEVEL_INFO) — quiet, production default.
  • 50 (DOCA_LOG_LEVEL_DEBUG) — useful debug detail.
  • 70 (DOCA_LOG_LEVEL_TRACE) — every call traced; the right setting for any first run of any new code path.

The trace build flavor enables additional sanitation that surfaces as log lines at DEBUG / TRACE level the release flavor would not emit.

Capability snapshots. doca_caps (CLI) and the library's own capability-query API (programmatic) are the source of truth for what this program can attempt on this host. Capture both at the start of any debugging session; a stale capability picture is a leading cause of "it worked yesterday".

Library-specific surfaces. Pipe counters, queue statistics, channel send / recv counters, and tracing dumps are owned by each library. The library skill names the call (e.g. doca_flow_pipe_query) and the right cadence to query it; this skill does not duplicate that.

Where logs go in long-running services. DOCA services (DTS, telemetry exporter, etc.) configure their own log paths. Check /var/log/doca* or the service's own README before assuming output is missing; do not assume stderr.

Safety policy

DOCA programs share state with the system, the device, and (in production) other tenants on the same fabric. The cross-library safety rules every library skill inherits:

  1. Validate before commit. Every library that programs hardware exposes a validate call separate from the commit / start / program call. Use validate first; never enter a commit path with an un-validated spec. Library skills extend this with library-specific what to validate (Flow validates pipe specs; RDMA validates queue-pair attributes; Comch validates channel cfg).
  2. Stage first, widen later. Run any new program against the smallest possible scope first — one representor, one queue, one channel. Read the output. Only after the staged run succeeds may the user widen the scope. The library's unit of damage (a representor for Flow, a QP for RDMA, a channel for Comch) defines what "smallest scope" means; the library skill names it.
  3. Trace flavor for development; release flavor for performance. The -trace build variant adds runtime sanitation that is invaluable while learning the API surface but introduces measurable overhead. Make the switch to release a deliberate, documented step before any performance reading is taken seriously.
  4. Quote the version on every API claim. Capability availability changes between releases. Statements of the form "this works in DOCA 3.3" must come from the user's installed pkg-config --modversion doca-<library>, not from agent memory or a public docs page that may describe a different release.
  5. Respect the env-side safety policy. Hugepage changes, eswitch mode changes, mlxconfig resets, BFB updates — all are global system state. The env-side rules in doca-setup CAPABILITIES.md ## Safety policy bind every program; no library override changes them.
  6. Never auto-modify the install tree. /opt/mellanox/doca/lib*/ is the release. Programs link against it; programs do not rewrite it. Build flavor changes go via LD_LIBRARY_PATH or the -trace pkg-config module, never by editing lib/.
  7. Trust the C ABI as the source of truth, even from non-C languages. Wrappers, FFI bindings, and codegen toolchains all eventually call the same libdoca_<library>.so the C samples link against. A binding-side decision that contradicts the public C ABI's lifecycle, error handling, or capability gate is wrong even if the binding compiles cleanly; the symptoms surface at runtime as DOCA_ERROR_BAD_STATE or DOCA_ERROR_INVALID_VALUE from the underlying C call.

Source: SKILL.md on GitHub

1 warning2mo3 checks · Risk SAFE
  • Gen Agent Trust Hub2mo

    The skill provides library-agnostic guidance for NVIDIA DOCA programming, including workflows for building, running, and debugging applications. It uses standard build tools and references official NVIDIA resources, with no security issues detected.

  • Socket2mo

    No alerts

  • Snyk2mo

    Risk: MEDIUM · 1 issue

Signed by skilld at f64fa0e. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub yesterday.

Activeupdated 3 months ago
metadata
{
  "kind": "library"
}
Other metadata
compatibility
No DOCA install required to read this skill (it is an overlay loaded against any DOCA artifact skill); the validation steps within DO require a live DOCA install at /opt/mellanox/doca.

README badge

README badge for nvidia/skills/doca-programming-guide