DOCA Documentation v3.3.0

Changes and New Features

  • NVIDIA ConnectX-9 SuperNIC Support – DOCA 3.3.0 introduces support for the NVIDIA ConnectX-9 SuperNIC, making it the first GA DOCA release to enable ConnectX-9

  • DOCA 3.3 is validated with the following NVIDIA platforms: GB200, GB300, B300, B200

  • DMS

    • Containerized DMS support

  • DOCA packages and components versioning

    • DOCA components version change only when component includes changes

    • OFED Packaging: Remove obsolete knem module and replace it with xpmem

  • DOCA Logging

    • doca-log users can control verbosity level per DOCA module

  • DOCA Host

    • Adding DOCA PCC to DOCA Networking Profile

    • Adding doca-sosreport to DOCA-Host Profiles (DOCA OFED, RoCE)

    • Removing SHARP from DOCA-Host

    • Supporting RISC-V CPU Architecture on DOCA-Host DOCA-OFED profile

    • Removing support for EOL: Ubuntu 20.04 , Ubuntu 25.04, Azure Linux 2.0, OpenEular 20.03, Debian 10.13, Xenserver 8.2

  • DOCA Flow

    • FlexParser [Beta] - allows acceleration of more complex/user-defined protocols

    • Allowing all IPv4/IPv6 encapsulation combinations (4 in 4, 6 in 6, 4 in 6, 6 in 4)

    • Enabling One-Way-Delay (OWD) calculation

  • DOCA Perftest

    • To run from DPU supporting bare-metal and N-S flows such as inference and storage

  • DOCA Platform Framework (DPF)

  • DTS

    • OVS Steering counters added to Multi-plane layer 2 solution

  • NVIDIA Network Operator

    • SPC-X K8s Support for Single-plane deployment for BlueField-3 SuperNIC [GA]

    • SPC-X K8s Support for software Multi-plane deployment for ConnectX-8 SuperNIC [GA]

    • For more information see NVIDIA Network Operator Release Notes

  • OVS-DOCA

    • OVS DOCA Added support for PTP link local traffic

    • OVS-DOCA upstream – basic acceleration support

    • IPv6 support improvements

    • Improved logging and troubleshooting for OVS DOCA

  • Platform

    • Expanded DPU scalability to support multiple switch ports and up to 32 PFs

    • Aligned BlueField attestation capabilities with the NVIDIA RIM Service, enabling automated access to CoRIM files to support zero-trust measurement and validation processes

    • Replaced MLNX_DPDK with official upstream DPDK (v25.11+) across the BlueField ecosystem, updating OVS and DOCA components to support standard upstream APIs.

    • Added support for software-defined vBLK emulation, utilizing the DOCA TLP SDK to offload storage data paths while retaining flexibility for custom control plane implementations

    • Added Redfish OOB support for configuring NIC Host Restriction settings, enabling administrators to enforce Zero Trust policies (equivalent to mlxprivhost) without logging into the DPU OS.

Note

NVIDIA® BlueField®-3 networking platforms must use DOCA-Host as the host driver. MLNX_OFED does not support BlueField-3 devices.

  • NVIDIA ConnectX-9 SuperNIC Support – DOCA 3.3.0 introduces DOCA-Host support for the NVIDIA ConnectX-9 SuperNIC, making it the first GA DOCA release to enable ConnectX-9.

  • RS-FEC Histogram Statistics Reporting – Adds support for reporting RS-FEC histogram statistics via ethtool. These IEEE 802.3–defined histograms reveal the distribution of corrected codewords by number of symbol errors corrected per block, improving link quality diagnostics and enabling earlier detection of signal degradation.

  • PTP Raw Cycle Counter Exposure – Extends the PTP subsystem and driver to expose the raw free-running cycle counter of PTP hardware clocks, enabling user space to obtain (cycle, system-time) pairs via new ioctls and correlate hardware cycle timestamps with host time without requiring PHC synchronization.

  • TCP Header-Data Split Zero-Copy – Exposes header–data split to the TCP layer to enable TCP zero-copy. This allows all headers up to and including TCP to be processed in software, while payload data is DMA-transferred directly from the NIC to the GPU without CPU involvement, achieving true zero-copy.

  • PCIe Congestion Event Counters – PCIe congestion events are events generated by the firmware when the device side has sustained PCIe inbound or outbound traffic above certain thresholds. This feature adds support for receiving and exposing such events as ethtool counters.

  • PH and Steering Tags Support – Enables the device to provide "hints" indicating where MR TLPs should be processed in system memory.

  • InfiniBand Service Record Resolution – Introduces APIs that translate an InfiniBand service name or ID into destination address details and allow querying that resolved IB addressing information.

  • IPsec Offload for Active-Passive Bonding – Active-passive bonding aggregates multiple network devices into a single logical interface for higher redundancy. While bonding already supported IPsec offload, this completes full support by adding correct IPsec GSO offload and fixing several race conditions around failover.

  • PSP Security Protocol – PSP Security Protocol support is now integrated into the core subsystem, with corresponding support implemented in the mlx5 driver. This enables mlx5 devices to recognize and handle PSP-secured traffic using the common core infrastructure.

  • Validation of DOCA-Host on ESXi Hypervisor – This feature validates that DOCA Host runs reliably in an ESXi hypervisor environment. It covers install/upgrade flows, correct device enumeration and passthrough configuration, and successful driver/module and service bring-up.

  • DOCA PCC inclusion in DOCA-Networking Profile – The DOCA-Host DOCA-Networking profile now includes DOCA-PCC, FlexIO, and DPACC.

  • DOCA-Host Universal doca-sosreport Enablement – Adds doca-sosreport to all DOCA-Host profile packages (RoCE, OFED, Networking, and All) so the same support data collection tool is always available, regardless of which DOCA flavor is installed. It standardizes how system and DOCA-related diagnostics are gathered (e.g., host OS details, driver/module status, configuration, and relevant logs), making it faster and more consistent to capture troubleshooting bundles for analysis and support.

  • Transition to Upstream DPDK Sourcing and Dependencies (DPDK 25.11+) – As DOCA transitions from MLNX_DPDK to upstream DPDK, all package dependencies switch to the upstream versions, and DPDK sources should be pulled from the official upstream repository, starting with DPDK 25.11 and tracking the latest releases from there.

  • Querying Effective VF Bandwidth when in LAG Mode – Currently, the mlx5 driver reports only the parent function's link speed to VFs, with no way to retrieve the VF's effective bandwidth in LAG and MPESW setups. As a result, user space and upper-layer software cannot obtain accurate bandwidth information, which can lead to suboptimal traffic decisions. This new capability allows user space applications to query the effective port speed and receive real-time notifications when it changes. In LAG configurations, each mlx5 port reports the aggregated transmit bandwidth available to it.

  • Extended DCB Traffic Class (TC) Rate Limiting - Extended DCB Traffic Class (TC) Rate Limiting, significantly increases the maximum configurable TC bandwidth on high-speed NICs. Previously, the TC max-rate field was limited to 8 bits, capping the value at 255 Gbps. With compatible firmware, the mlx5 driver now supports TC rate limits up to 65,535 Gbps (≈ 6.5 Tbps) through capability detection, while maintaining backward compatibility on legacy hardware.

    Note

    The configured maximum must align with the limits supported by the specific hardware.

    Example configuration methods:

    • DCB utility

      Copy
      Copied!
                  

      dcb maxrate set dev eth2 tc-maxrate 0:0 1:0 2:0 3:0 4:0 5:0 6:0 7:300Gbit

    • mlnx_qos

      Copy
      Copied!
                  

      mlnx_qos -i eth2 -r 5,0,0,10,0,0,7,320

  • PCIe Attention Button Hotplug Support (virtio-net/virtio-blk) – Added a new firmware capability that implements the PCIe-specification attention button hotplug/unplug flow for virtio-net and virtio-blk. This enables standards-compliant user-initiated device insertion/removal signaling and proper handling by the host stack for these virtio functions.

  • DOCA PCC – The DOCA PCC NP application now enables the NIC to insert the RTT response transmit timestamp in hardware, reducing software-induced jitter and improving the accuracy and consistency of RTT measurements.

  • Fail-Safe DPU Identification via NC-SI – Implemented a new NC-SI MLNX command, get_host_info, that allows software to reliably identify which DPU corresponds to a given physical machine/PCIe slot. This provides a fail-safe identification mechanism for inventory, provisioning, and troubleshooting workflows, especially in systems with multiple DPUs.

  • DPA Process Limit Update – The system-wide limit for DPA processes has been reduced to 30. This total includes both user processes across all GVMIs and internal ProgCC processes. The max_dpa_processes value reported to the user is calculated as: max_dpa_processes=30−number_of_progcc_processes

  • PLDM PDR Repository Change Event Support – PLDM now supports the PDR Repository Change event type, enabling notification to the BMC when PDRs change. With this flow, the BMC can detect cable insertion/removal events. Refer to DSP0248 for details.

  • DOCA PCC API: DSCP Query for PCC Flows in QP Mode – Introduces a DOCA PCC device API that enables retrieval of the DSCP value associated with a PCC flow when PCC operates in QP mode (for example, when ROCE_CC_SHAPER_COALESCE_P1=2 and ROCE_CC_SHAPER_COALESCE_P2=2 are configured via mlxconfig).

  • BlueField-3 DPU Mode VF Migration Support – Introduced support for migrating Virtual Functions (VFs) in BlueField-3 DPU mode. It enables a VF’s state and connectivity to be preserved across migration events, improving continuity for virtualized workloads and reducing disruption during maintenance operations such as host servicing, live migration, or recovery workflows.

  • Parallel Save/Load Support for VF Migration – Added support for running save and load operations in parallel, enabling multiple contexts (e.g., multiple VFs) to be checkpointed and restored concurrently instead of serially. This reduces overall migration time and improves scalability in environments that need to migrate or recover many VFs at once.

  • NVGRE VSID Modify-Header Support - Extended packet modify-header operations to support set and copy actions on the NVGRE VSID (Virtual Subnet Identifier). A new field, TUNNEL_HDR_DW_2 (0x84), enables dynamic VSID modification, adding header rewrite support for NVGRE tunnel traffic in addition to existing filtering capabilities.

  • Advanced SNAP NVMe Hot-Plug/Unplug Flow - Advanced SNAP NVMe Hot-Plug/Unplug Flow introduces a recommended two-step flow for hot-plugging and unplugging emulated PCI PF devices. The device is first created in a POWER_OFF state, fully configured by DPU software, and only then exposed to the host PCI subsystem. This provides better control and visibility, makes the plug/unplug flows more symmetric, and helps avoid undefined PCI states.

  • SNAP NVMe Host-Aware/Host-Unaware Hot-Plug Mode - Host awareness ensures the DPU tracks the host system state (e.g., boot, power-off, or power-cycle) and performs hotplug/hotunplug operations only when the host can safely handle PCIe changes. In host-aware mode, NVMe function hotplug/unplug explicitly gates these actions on host readiness; in host-unaware mode, these checks are skipped, which is unsafe and may lead to unpredictable behavior.

  • PLDM PDR Repository Change Event Support – PLDM now supports the PDR Repository Change event type, enabling notification to the BMC when PDRs change. With this flow, the BMC can detect cable insertion/removal events. Refer to DSP0248 for details.

The following table details the API changes that were introduced in this version of DOCA:

Library

Change Description

TBD

TBD

Info

To view API changes for previous DOCA versions, refer to the release notes of those archived versions.

© Copyright 2026, NVIDIA. Last updated on Mar 2, 2026