Doca Gpunetio

DOCA GPUNetIO workflows

Where to start: The verbs run configure → build → modify → run → test → debug. Skip ahead only when the user is already past a verb. The ## test verb is an iterative loop (single-packet smoke → persistent-kernel burst loop → multi-GPU scale-out → loop back if a precondition changes), not a one-shot pass — see the eval-loop overlay in ## test below.

Read this file when the loader sent you here from SKILL.md. For the GPUNetIO capability surface, the doca_gpu per-device context, the GPU-visible queue handles, the persistent-kernel pattern, the dual capability-query rule, the env-precondition matrix, the error taxonomy, the observability surface, and the safety policy, see CAPABILITIES.md. For the cross-library DOCA patterns layered under everything below (the universal lifecycle, the cross-library DOCA_ERROR_* taxonomy, the modify-a-shipped- sample workflow), see doca-programming-guide/SKILL.md.

Each verb below describes the shape of the workflow, not a copy-paste recipe. The agent's job is to walk the user through the steps in order, verifying preconditions before recommending the next call.

DOCA 3.5 field notes

GPU memory registration: dmabuf first, nvidia-peermem only as a fallback (DOCA 3.5 GPUNetIO installation and setup page). dmabuf is the default and preferred method; it needs Linux kernel 6.2 or later, libibverbs 1.14.44 or later, and the open NVIDIA kernel driver (nvidia-open; CUDA 12.5 or older must be installed with -m=kernel-open). Get the descriptor with doca_gpu_dmabuf_fd(); on success use doca_mmap_set_dmabuf_memrange(), otherwise doca_mmap_set_memrange(). Without dmabuf, doca_mmap_start() falls back to legacy registration through nvidia-peermem (sudo modprobe nvidia-peermem) and logs Registration using dmabuf is not supported, falling back to legacy registration, which is harmless if peermem is acceptable. A missing nvidia-peermem module is therefore only a problem when dmabuf is unavailable.

DOCA 3.5 public API changes (public GPUNetIO header, doca_gpunetio.h). Update code that uses the GPUNetIO Verbs API:

• Stop reading QP/CQ fields directly from struct doca_gpu_verbs_qp. Use the new doca_gpu_verbs_get_qp_* getter family to obtain the underlying QP, send CQ, and receive CQ.

• Use doca_gpu_verbs_req_notify_cq() for CQ notification instead of the previous notification path.

• Use doca_gpu_nic_handler_to_string() when a NIC handler name is needed.

• Update constant usage to the DOCA 3.5 DOCA_GPUNETIO_VERBS_* and DOCA_GPUNETIO_ETH_* families from the public header.

See the public GPUNetIO header for the exact declarations and constant families.

configure

Goal: stand up a doca_gpu context on a target CUDA device, layer GPU-visible queue handles on top of an already-configured doca-eth queue pair, and confirm both the DOCA side and the CUDA side are in a state where a persistent kernel can be launched.

Steps the agent should walk the user through:

1. Confirm the env preconditions. Per the env-precondition matrix in CAPABILITIES.md ## Safety policy: nvcc --version returns a CUDA toolkit version listed as compatible with the installed DOCA release by the DOCA Compatibility Policy; GPU memory registration is available — dmabuf first (see DOCA 3.5 field notes below), and only when dmabuf is unavailable lsmod | grep nvidia_peermem shows the fallback module loaded (or sudo modprobe nvidia_peermem); nvidia-smi -L enumerates the candidate GPU; the underlying doca-eth queues exist (route to doca-eth/SKILL.md for the upstream verb). If ANY of these fails, this is an env / version problem to fix via doca-setup TASKS.md ## configure (doca-setup/TASKS.md#configure)

   • doca-version TASKS.md ## configure (doca-version/TASKS.md#configure), NOT a code change in the GPUNetIO program.

2. Bring up doca-eth FIRST. GPUNetIO sits on top of doca-eth: the user must have a configured doca_eth_rxq (and doca_eth_txq if TX is in scope) on the target representor / PCIe device, with the doca-eth context at the right lifecycle stage for the GPUNetIO layer to attach. Per the layering rule in CAPABILITIES.md ## Capabilities and modes, the agent must NOT skip ahead to doca_gpu_* calls if doca-eth is not already up — that ordering is fixed for every version of GPUNetIO.

3. Run dual capability discovery against the target device. Per the dual-axis rule in CAPABILITIES.md ## Capabilities and modes: on the DOCA side, run the doca-eth query doca_eth_rxq_cap_is_type_supported(devinfo, ...) from doca_eth_rxq.h (and the matching doca_eth_txq_cap_is_type_supported / doca_eth_*_cap_get_* sizing family) against the active doca_devinfo for the NIC — GPUNetIO exposes no cap_is_supported symbol of its own; on the CUDA side, run cudaGetDeviceProperties(devOrdinal) against the candidate CUDA device. Quote BOTH results back to the user. If either says not supported, that axis is the answer — do not proceed.

4. Allocate the GPU payload pool and register it with DOCA. The receive payload pool lives in GPU memory: allocate via cudaMalloc (or the CUDA allocator the user's app already uses) on the target device, then register the buffer with DOCA via the doca_buf_arr_create_* family before starting the parent doca-eth context with doca_ctx_start(). Per the safety policy in CAPABILITIES.md ## Safety policy, a missing or out-of-order registration surfaces as DOCA_ERROR_BAD_STATE from the first device-side submit, not at create time.

5. Create the doca_gpu context and the GPU-visible queue handles. Create one doca_gpu for the target GPU — doca_gpu_create(const char *gpu_bus_id, ...) takes the GPU's PCIe bus-id STRING (e.g. from nvidia-smi -L / cudaDeviceGetPCIBusId), not an integer device ordinal; then create the GPU-visible RX handle (doca_gpu_eth_rxq) on top of the existing doca_eth_rxq (and, if TX is in scope, the matching doca_gpu_eth_txq). Start the doca-eth context if it has not been started already. Per the lifecycle ordering in CAPABILITIES.md ## Safety policy, record the order so teardown happens in reverse: GPU-visible handle → doca_gpu → CUDA context → doca-eth queue.

6. Sanity check before launching any CUDA kernel. Confirm with the user: which CUDA device ordinal the persistent kernel will run on; which doca-eth queue the GPU-visible handle is layered on; whether RX-only, TX-only, or both; how the persistent kernel will know when to exit (a host-side flag read from device memory is the canonical signal). If any of those are unclear, stop and ask — do not invent.

For the canonical DOCA universal lifecycle that underlies steps 2-5, see doca-programming-guide TASKS.md ## configure (doca-programming-guide/TASKS.md#configure). This skill adds the GPUNetIO overlay; do not re-explain the lifecycle here.

build

Goal: compile a GPUNetIO-using consumer (host-side C/C++ + at least one CUDA .cu translation unit) against the user's installed DOCA + CUDA toolkit, with pkg-config + nvcc as the joint sources of truth for include + link flags.

The build pattern for any DOCA C/C++ consumer is fully documented in doca-programming-guide TASKS.md ## build (doca-programming-guide/TASKS.md#build). GPUNetIO adds a CUDA layer on top: the device-side code lives in .cu files compiled by nvcc, and the link line pulls in both the DOCA libraries and the CUDA runtime. This skill carries only the GPUNetIO-specific overlay:

For non-C host-side consumers (Rust, Go, Python) that drive GPUNetIO setup and launch a CUDA kernel built separately, the host-side link line and version rules above still apply; the CUDA kernel itself is a separate compilation unit and is out of scope for this skill, but the nvcc version check is the load-bearing input the wrapper still needs.

modify

Goal: take the closest-fitting shipped DOCA GPUNetIO sample as the verified starting point and apply a minimum diff to make it match the user's intent, without rewriting from scratch.

The universal modify-a-shipped-sample workflow is in doca-programming-guide TASKS.md ## modify (doca-programming-guide/TASKS.md#modify); this skill provides the GPUNetIO-specific slot fill.

The agent emits an intent description + the six filled slots; the actual unified diff against the sample source is produced the way every other library skill in this bundle handles modify — the agent walks the user through the diff line-by-line against the sample source they read on disk, and has the user paste back the result for validation. The agent's anti-pattern alert: a "clean rewrite" from scratch is almost always slower to first green than a minimum-diff modify on a shipped GPUNetIO sample, and removes the user's ability to bisect against a known-good baseline.

run

Goal: actually launch the persistent kernel against the configured GPU-visible queue and confirm packets flow on the wire.

Steps the agent should walk the user through:

1. Confirm the peer is generating traffic. GPUNetIO needs packets arriving at the doca-eth RX queue; running the persistent kernel against an empty queue produces a misleading 0% GPU util and looks like a hang. The agent must surface this on every first run.

2. Launch the persistent kernel with the GPU-visible RX (and optionally TX) handle passed in as a kernel argument. The kernel is launched once and runs for the duration of the data-movement session; do not re-launch per batch. The block / thread sizing depends on the user's GPU per cudaGetDeviceProperties.

3. Drive the host-side doca_pe_progress loop in parallel with the persistent kernel — the DOCA-side completions and the connection / lifecycle events still flow through the host PE even when the data plane is GPU-driven. A common silent-failure mode is forgetting the host-side PE drive and wondering why state transitions are not reported.

4. Capture the structured log on first failure. Set DOCA_LOG_LEVEL=trace for the first run (see doca-debug CAPABILITIES.md ## Observability (doca-debug/CAPABILITIES.md#observability)); pair it with nvidia-smi -q -d UTILIZATION while the kernel runs to confirm the GPU is actually busy. A persistent kernel that reports 0% util is stuck on a queue drain.

For the runtime version + LD_LIBRARY_PATH cross-checks that underlie "the program built but does nothing", see doca-version TASKS.md ## run (doca-version/TASKS.md#run).

test

Goal: prove the configured GPUNetIO setup actually moves packets from the wire into the persistent kernel on the user's hardware, and that the env-precondition + capability set was sized right.

This is a loop, not a one-shot pass. Each iteration narrows either the env-precondition set, the capability set, the persistent-kernel design, or the buffer sizing. Termination and the iteration-identity tuple are defined once under Loop termination below.

Iteration shape:

1. Single-packet smoke from the GPU side. Send ONE packet from a peer; confirm the persistent kernel observes one RX completion. If yes, advance. If no — and nvidia-smi reports the kernel is running — the GPU-visible RX handle is not actually layered on the doca-eth queue that the peer's traffic is hitting; re-run ## configure step 5.

2. Persistent-kernel burst loop. Send 1000 packets in a burst; confirm the persistent kernel processes all of them without dropping. Catches buffer-pool-sizing bugs: DOCA_ERROR_AGAIN from the device-side submit means the queue is full and the persistent kernel needs to drain faster, OR the buffer pool is undersized — per ## modify slot 5.

3. TX smoke (if TX is in scope). Submit one TX packet from the persistent kernel via the GPU-visible TX handle; confirm it arrives at the peer. The fast-path TX submit from device code is independent of the RX path and can fail differently (different cap-query axis).

4. Multi-GPU scale-out (if used). Re-run steps 1-3 on each doca_gpu independently — different GPUs in the same host can fail the dual capability check differently (e.g. an Ampere card supports GPUNetIO while a Turing card in the same chassis does not). Per CAPABILITIES.md ## Capabilities and modes the per-GPU rule, each card is its own configuration.

5. Negative test. Intentionally point the persistent kernel at a CUDA device that should NOT support GPUNetIO (e.g. a pre-Ampere card) and confirm the dual capability discovery from ## configure step 3 reports the DOCA_ERROR_NOT_SUPPORTED cleanly — validates the agent's capability-discovery itself is correct.

Eval-loop overlay — why this is a loop, not a one-shot pass:

Loop termination: the loop ends when either (a) the persistent kernel processes packets at the user's intended rate with the expected accuracy — success — or (b) two consecutive iterations produce the same iteration-identity tuple (trigger-table row, doca-gpunetio / doca-common / CUDA version results, DOCA and CUDA capability results, parent doca-eth queue identity and counter outcome, persistent-kernel state, and packet / drain outcome). Matching evidence across both iterations requires escalation, never success; the cause is below GPUNetIO (CUDA driver, NIC firmware, network). Escalate to doca-debug TASKS.md ## debug (doca-debug/TASKS.md#debug) with the captured layer-1-through-5 evidence including BOTH the DOCA log and the nvidia-smi output.

debug

Goal: when a DOCA GPUNetIO call (or the persistent kernel) returns a DOCA_ERROR_* or does not make forward progress, narrow the cause to a specific layer and act on it.

The cross-library debug ladder lives in doca-debug TASKS.md ## debug (doca-debug/TASKS.md#debug). Walk through it in order — install → version → build → link → runtime → program → driver — before recommending GPUNetIO-specific fixes. This skill's overlay names the GPUNetIO-specific manifestation at layers 5 (runtime), 6 (program), and 7 (driver):

Layer 5 (runtime) — GPUNetIO overlay.

• The persistent kernel never observing a completion is almost always one of three things: the underlying doca-eth queue is not actually receiving traffic; the GPU-visible handle is layered on a different queue than the one the peer is hitting; or the host-side PE is not being progressed. Confirm the env-side preconditions per ## configure step 1 before any code change.

• DOCA_ERROR_AGAIN from device-side submit is always a queue-full / drain-rate problem. Do not recommend a retry loop in the kernel; recommend a faster drain step in the persistent kernel body.

• A "persistent kernel hung; cannot stop the program" pattern is almost always a missing host-side termination flag per ## modify slot 6.

Layer 6 (program) — GPUNetIO overlay.

• Lifecycle ordering: the GPU-visible handle must be destroyed BEFORE the underlying doca-eth queue is destroyed; the doca_gpu context must be destroyed BEFORE the CUDA context is unbound from the process. Out-of-order returns DOCA_ERROR_BAD_STATE per CAPABILITIES.md ## Safety policy.

• Buffer registration: a CUDA-allocated buffer that was not registered with DOCA via doca_buf_arr_create_* before doca_ctx_start() returns DOCA_ERROR_BAD_STATE from the first device-side submit, not at create time. Walk the registration sequence in ## configure step 4.

• Kernel-per-packet anti-pattern: if the user reports "throughput is much lower than expected" and the code launches a kernel per packet, that is the bug. Refactor to the persistent-kernel pattern per CAPABILITIES.md ## Capabilities and modes.

Layer 7 (driver) — GPUNetIO overlay.

• DOCA_ERROR_DRIVER from doca_gpu_* is most often a CUDA driver layer reporting failure to DOCA. Capture nvidia-smi -q and dmesg | tail; cross-check the CUDA driver version against the DOCA Compatibility Policy.

• nvidia_peermem not loaded silently degrades GPUDirect RDMA and surfaces as DOCA_ERROR_NOT_SUPPORTED at create time. Route to doca-setup TASKS.md ## debug (doca-setup/TASKS.md#debug) layer 5 (driver) for the kernel-module fix.

Once the layer is identified, route to the matching debug verb on the matching skill: install / build / link / driver to doca-setup ## debug (doca-setup/TASKS.md#debug); version to doca-version TASKS.md ## debug (doca-version/TASKS.md#debug); cross-cutting runtime to doca-debug ## debug (doca-debug/TASKS.md#debug); program-layer Core-context patterns to doca-programming-guide TASKS.md ## debug (doca-programming-guide/TASKS.md#debug).

5-phase universal debug-loop instantiation (GPUNetIO). Layer identification above is phase 1 of the universal debug-loop contract (doca-debug/CAPABILITIES.md#universal-debug-loop-contract). The agent MUST walk the remaining four phases on every GPUNetIO debug answer before declaring done:

1. Layer identification — above (RX / lifecycle / driver).

2. Triple capture (READ-ONLY). Capture (a) the saved configure-time doca_devinfo, parent doca_eth_rxq identity, doca_eth_rxq_cap_is_type_supported(...) result, and ethtool -S <netdev> RX counters, (b) DOCA log lines at DOCA_LOG_LEVEL=DEBUG for the offending submit / drain, and (c) GPU-side state via nvidia-smi -q + cuda-memcheck (or compute-sanitizer) on the persistent kernel. Capture ALL THREE before mutating; the triple is the rollback target.

3. Single-variable mutation SMALLER than the original change. Examples: drop the kernel's per-iteration drain width by half (not refactor to per-packet); register one extra buffer (not the full pool); switch one queue to the doca-eth smoke sample (not the production topology). Larger mutations void the experiment.

4. Re-capture and compare. Re-run the triple from phase 2; diff against the baseline. The compare diff IS the evidence, not the agent's prose interpretation.

5. Exit with named green signal OR escalate. Green = the specific counter / log line / GPU state the bug was negating. Examples: ethtool -S rx_packets increments AND GPU buffer holds the bytes; no compute-sanitizer errors; stop-flag termination clean. If two consecutive iterations don't change anything, the cause is below GPUNetIO — escalate to doca-debug TASKS.md ## debug (doca-debug/TASKS.md#debug) with the captured triple.

rollback

GPUNetIO contexts are stateful (persistent kernel + GPU buffer registration + doca-eth queue) and the agent's failure mode is to leave half-registered GPU buffers behind after a botched configure → start sequence. The universal verification contract (doca-setup/CAPABILITIES.md#universal-verification-contract) step 1 (preconditions) requires "the rollback path is documented" on every change-recommending answer; this is the GPUNetIO instantiation.

Snapshot before mutate. Before any change-recommending GPUNetIO answer, capture the GPU-side allocation map (nvidia-smi -q -d MEMORY + cudaGetDeviceProperties), the doca-eth queue attachment map (the saved configure-time doca_devinfo, eth RXQ identity, and doca_eth_rxq_cap_is_type_supported(...) result), and the persistent-kernel stop-flag location. The triple IS the rollback target.

1. Signal the persistent kernel to drain and stop FIRST. Flip the host-side termination flag from ## modify slot 6 before any context destroy. Before flipping it, record a fixed drain deadline. Poll cudaStreamQuery on the kernel's stream until it reports completion or that predeclared deadline expires; do not call cudaDeviceSynchronize for this drain. If the deadline expires, the kernel is hung — that is the deploy-loop bridge (doca-setup/CAPABILITIES.md#deploy-loop-bridge--step-5-not-green-is-the-debug-loop-trigger) trigger, not a rollback trigger; fire the debug-loop on the hung-kernel symptom. Fail closed: do not continue to steps 2-4 or call doca_buf_arr_destroy, doca_ctx_stop, doca_gpu_destroy, or cudaFree while kernel completion is unconfirmed. Preserve the rollback snapshot and escalate with the captured debug triple. Resume rollback only after cudaStreamQuery reports completion.

2. Unregister GPU buffers in reverse-register order. doca_buf_arr_destroy on every array created with doca_buf_arr_create_*. Buffers MUST be unregistered before cudaFree; the reverse order is non-negotiable per CAPABILITIES.md ## Safety policy.

3. Stop and destroy the GPUNetIO context. doca_ctx_stop on the doca_gpu context, then doca_gpu_destroy. The parent doca-eth RXQ remains valid and must not be torn down by this step.

4. cudaFree the underlying GPU allocations. Only after the DOCA-side unregister landed clean. Skipping this step leaks GPU memory at process exit and surfaces as CUDA_ERROR_LAUNCH_FAILED on the next program run with the same NUMA-pinned GPU.

5. Re-verify the doca-eth parent is intact. Re-run the eth RXQ smoke from doca-eth TASKS.md ## test (doca-eth/TASKS.md#test) to confirm the parent queue still receives packets at the pre-GPUNetIO rate. If not, the GPUNetIO rollback corrupted the parent — that is a bug to surface, not a retry trigger.

Verification-contract requirement. Document the rollback verb in the verification contract preconditions block. The step 1 line for a GPUNetIO add reads: "the rollback path is the five-step reversal in ## rollback; the agent has captured the GPU allocation map and persistent-kernel stop-flag location." Without that line, the contract is incomplete and the agent is NOT eligible to declare done.

The rollback is bounded — on the second non-green re-verify at step 5, the agent MUST surface the unresolved residual gap instead of recommending another GPUNetIO retry.

Deferred task verbs

The following verbs are out of scope for this skill but are commonly asked in the same conversations. Route them as follows so the agent does not invent guidance:

• install. Installing DOCA, installing the CUDA toolkit, choosing matched versions, post-install verification, pkg-config wiring, loading nvidia_peermem — defer to doca-setup/SKILL.md and to the install-tree layout in doca-public-knowledge-map ## Layout of an installed DOCA package (doca-public-knowledge-map/SKILL.md#layout-of-an-installed-doca-package). This skill assumes DOCA + CUDA are already installed and matched.

• deploy. Deploying GPUNetIO-using applications at scale (multi-GPU multi-NIC clusters, Kubernetes operator workflows for GPU + DPU workloads, multi-tenant GPU sharing for GPU-initiated networking) — out of scope for this skill. For single-host first-run testing, the right verb in this skill is ## run; do not invent a "deploy" workflow.

• CUDA kernel design and tuning. Block / thread sizing, warp-level optimization, shared memory layout inside the persistent kernel — out of scope. Route to the upstream CUDA toolkit documentation via doca-public-knowledge-map/SKILL.md; this skill prescribes that the user uses the persistent-kernel pattern, not how to tune it.

• DOCA Ethernet queue setup. Bringing up the underlying doca_eth_rxq / doca_eth_txq, RSS configuration, representor selection — GPUNetIO depends on it but does not redefine it. Route to doca-eth/SKILL.md.

Command appendix

Every command below is cross-cutting on DOCA GPUNetIO — it answers a recurring class of question that comes up in the verbs above. The agent should treat the class as load-bearing; the worked example is a single instance. Run-as user is the unprivileged user unless noted. Rows that need elevated privileges call that out explicitly.

Infra-aware preamble (every row below). Per the bundle's detect → prefer → fall back → report contract documented in doca-structured-tools-contract ## The agent behavior contract (doca-structured-tools-contract/SKILL.md#the-agent-behavior-contract), the agent should:

1. Probe for the matching structured helper FIRST (doca-env --json for version + devices + libraries + drivers + hugepages in one shot; doca-capability-snapshot for per-device capability flags; version-matrix.json for "available since" lookups).

2. If the probe succeeds, the structured tool's output is the authoritative answer and the agent SHOULD NOT also run the manual command in the row below. Report "using structured <tool>".

3. If the probe fails, fall back to the manual command in the row. Report "falling back to manual chain".

4. The schemas the structured tools emit are defined in doca-structured-tools-contract ## Schemas (doca-structured-tools-contract/SKILL.md#schemas); the version-handling semantics (four-way match, NGC, headers-win) are owned by doca-version/SKILL.md.

For commands shared across libraries (pkg-config --modversion, doca_caps, cat /opt/mellanox/doca/applications/VERSION, DOCA_LOG_LEVEL) the cross-library overlay is in doca-debug TASKS.md ## Command appendix (doca-debug/TASKS.md#command-appendix); this table adds the GPUNetIO-specific rows on top.