Shared-memory DMA and authenticated object transfer for DPDK in AMD SEV-SNP VMs.
CVM-DPDK extends DPDK 25.11.3's mlx5 driver and testpmd so a passed-through NIC can use explicitly shared guest memory for DMA while packet processing and authenticated object reassembly stay in private guest memory.
This is a research prototype, validated with an NVIDIA ConnectX-7 on a 100 Gb/s link inside an AMD SEV-SNP guest.
- Shared DMA allocation: a Linux
cvm_sharedmodule exports shared pages as dma-bufs; mlx5 patches register packet buffers and device resources. - Private packet processing: validated private CQE snapshots and a testpmd bridge keep application buffers and mbuf metadata in private memory.
- Authenticated object transfer:
cvmchunkadds AES-256-GCM chunks, authenticated ACKs, retransmission and private object reassembly.
Shared DMA memory remains untrusted. The bridge copies packets into private
memory; cvmchunk authenticates data before committing it to an object.
Simplified mlx5 RX flow for scalar SPRQ with CQE compression disabled. Blue = CPU access; orange = NIC DMA.
The CPU posts receive buffers through WQEs and updates the RQ doorbell record
(1–2). The NIC reads those requests, writes packet data and reports completion
(3–5); the CPU polls the CQ (6). CVM-DPDK places these DMA resources in shared
guest memory, then validates private CQE snapshots and copies packets into
private buffers. The private bridge and cvmchunk authentication are not shown.
How to read the diagram: scope and simplifications
- The four memory blocks cover the illustrated RX path, not every NIC DMA resource. TX and CQ consumer updates are omitted. “No interrupts” describes packet polling here, not all device events.
- The RQ doorbell record holds the posted-WQE producer counter. Together with its consumption progress, the NIC determines available receive requests; this is not the mempool's free-buffer count. Posting to this RQ requires a record update, not a per-packet MMIO write to the pictured NIC register.
- One CQE per successfully received packet describes the uncompressed path. The drawing uses 64-byte CQEs and example buffer/packet sizes; timestamps require the corresponding configuration.
- An
lkeyidentifies a registered memory region. Address translation also uses the requested address and length, with bounds and permission checks;lkeyalone is not a physical address.
See the mlx5 driver guide and hardware programming manual, §§6.1 and 7.4.
Run the portable checks on Debian/Ubuntu with Git 2.43 or newer:
git clone https://github.com/sept-usc/cvm-dpdk.git
cd cvm-dpdk
sudo apt-get install build-essential pkg-config libssl-dev python3 python3-cryptography
make test
make checkThese checks need no DPDK installation, NIC or SNP guest. To build the full path, follow the DPDK patch guide and hardware setup and reproduction guide.
Hardware tests used AMD EPYC 9354, a SEV-SNP guest at VMPL0, and a passed-through ConnectX-7 PF. Exact kernels, firmware and runtime settings are in the tested configuration.
| Test | Recorded result |
|---|---|
| Shared memory and NIC DMA | Six module checks and a byte-level macswap round trip passed |
| Production-code regressions | 48 checks covering CQE mutation, buffer ownership and forwarding passed |
| Upstream test status | 44 DPDK commands across native, ARM64, PCI virtio and vDPA profiles pass in CI; the ledger tracks later results and remaining prerequisites against the original 246-entry audit |
| Raw receive, before the private bridge | 40 GiB synthetic stream: 94.680 Gbit/s UDP payload, no authentication |
| Authenticated private-object transfer | 512 MiB synthetic object: 3.596 s / 1.194 Gbit/s, exact SHA-256, zero retransmissions; one queue/core |
These measurements cover different data paths and are not a controlled performance comparison. See the validation index for workloads, loss counters, negative controls and unresolved issues.
The CVM path currently supports scalar SPRQ and testpmd's private-buffer bridge. Production key management and replay protection across restarts are unfinished. GPU experiments are separate; authenticated delivery into GPU memory is not implemented. See configuration limits.
Open an issue for questions or bugs. See CONTRIBUTING.md for development and local Git hooks, and SECURITY.md for private vulnerability reports.
Original project material uses MIT. DPDK patches and protocol code retain BSD-3-Clause; kernel components retain GPL-2.0. See LICENSING.md for component boundaries.