What problem does it solve?
Developers writing RDMA data-movement applications with the NVIDIA DOCA RDMA library face a steep learning curve: eleven task types, three connection methods, an asymmetric mmap/RDMA permission matrix, capability queries that vary by device and DOCA version, and cryptic DOCA_ERROR_* failures. This Skill gives an AI agent the verified workflows and safety rules to configure, build, modify, run, test, and debug DOCA RDMA programs correctly instead of hallucinating APIs or falling back to raw libibverbs.
Core Features & Use Cases
- Lifecycle and connection guidance: Walks through bringing up an RDMA context on a doca_dev and choosing between RDMA CM, bridge/OOB, or gRPC export-based connection methods.
- Task and permission matrix: Covers all eleven task types (Send/Receive, Read/Write, Atomics, Sync Events) with the exact mmap + RDMA permission pairs and export requirements for each.
- Capability and version discovery: Enforces doca_rdma_cap_* queries and pkg-config version checks so recommendations match the actual device and installed DOCA release.
- Layered debugging: Maps DOCA_ERROR_* families (BAD_STATE, NOT_PERMITTED, NOT_SUPPORTED, FULL, DRIVER) to specific layers with an iterative test loop.
- Use Case: A developer hits "one-sided read returns permission denied" on a BlueField DPU. The agent loads this Skill, checks the permission matrix, confirms the peer mmap was exported with RDMA-read permission, and walks the user through a Send/Receive smoke test to isolate the failure.
Quick Start
Ask your agent to help you bring up a DOCA RDMA send/receive connection between two processes on your BlueField DPU and debug any DOCA_ERROR_* that appears.