Networking DOCA SDK Documentation

Bug Fixes

DOCA Framework Bug Fixes

Ref #

Issue Details

5357609

Description: When GDRCopy is active, doca_gpu_semaphore_start() might fail with DOCA_ERROR_DRIVER for a semaphore that uses the DOCA_GPU_MEM_TYPE_GPU_CPU memory type.

Keyword: DOCA GPUNetIO; semaphore; GDRCopy

Detected in version: 3.6.0

5357606

Description: In a process that has not created a CUDA context, doca_gpu_create() might return DOCA_SUCCESS without creating a GPU handle.

Keyword: DOCA GPUNetIO; doca_gpu_create; CUDA context

Detected in version: 3.6.0

5350951

Description: On a correctly configured host, DMS blueprint verification might report a huge-pages mismatch.

Keyword: DMS; DO-SPCX; verification; huge pages

Detected in version: 3.5.0

5319232

Description: On B300 systems that use the SPX_Multiplane profile, the DO-SPCX configuration stage might fail while setting up virtual functions.

Keyword: DMS; DO-SPCX; B300; SPX_Multiplane; virtual functions

Detected in version: 3.5.0

5309524

Description: OVS (ovs-vswitchd) might occasionally crash with a segmentation fault when a VF representor port is removed from an OVS-DOCA bridge.

Keyword: OVS-DOCA; DOCA Flow; VF representor; segmentation fault

Detected in version: 3.2.2

5303745

Description: Restarting doca_prdma_provider after a crash, while DOCA Lifeguard holds its RDMA resources, might fail with an EREMOTEIO error during hardware steering resource allocation.

Keyword: PRDMA; DOCA Lifeguard; process recovery; hardware steering

Detected in version: 3.6.0

5301395

Description: The DO-SPCX configuration stage might fail to start DOCA XPlane because the DMS launcher requires a container image tag, while XPlane is installed as native binaries from the doca-xplane package.

Keyword: DMS; DO-SPCX; DOCA XPlane; configuration

Detected in version: 3.6.0

5301165

Description: After an NVIDIA® BlueField®-3 is set to NIC mode and the host is power cycled, dmsc might still report the current operating mode as DPU.

Keyword: DMS; dmsc; NIC mode; operating mode

Detected in version: 3.5.0

5298787

Description: On GB300 systems, planning the DO-SPCX configuration stage might fail with a topology mismatch error that reports one plane PF per rail instead of two.

Keyword: DMS; DO-SPCX; GB300; planning

Detected in version: 3.6.0

5298723

Description: Restarting OVS with an OVS-DOCA bridge attached might occasionally cause OVS to hang because two PMD threads receive the same offload thread ID.

Keyword: OVS-DOCA; PMD threads; restart; hang

Detected in version: 3.6.0

5286143

Description: In MRC deployments on NVIDIA® ConnectX®-8, PS Manager (PSMA) might fail to start because it is launched with --pss-served-ips together with --repr or --vf-netdev-names, a combination it rejects.

Keyword: PS Manager; MRC; ConnectX-8; startup

Detected in version: 3.6.0

5285558

Description: On aarch64 hosts, the DO-SPCX boot policy might apply x86-only IOMMU kernel parameters, so IOMMU passthrough is not configured. On PXE- or HTTP-booted aarch64 hosts, the preparation stage might fail after every reboot.

Keyword: DMS; DO-SPCX; aarch64; IOMMU; boot policy

Detected in version: 3.3.1

5278009

Description: DO-SPCX might write the network and RDMA device naming rules to a single udev rules file instead of the two files that the deployment guide specifies.

Keyword: DMS; DO-SPCX; udev rules; interface naming

Detected in version: 3.3.1

5271557

5303130

Description: DO-SPCX might set cc_probe_mp_mode on software multi-plane and single-plane profiles, although only hardware multi-plane deployments require it. On single-plane deployments with ConnectX-7, this setting might stop all traffic.

Keyword: DMS; DO-SPCX; cc_probe_mp_mode; single-plane

Detected in version: 3.5.0

5269600

Description: On clusters without GPUs, such as CPU-only clusters, the DO-SPCX preparation stage might fail at the GPUDirect RDMA check.

Keyword: DMS; DO-SPCX; CPU cluster; GPUDirect RDMA

Detected in version: 3.5.0

5269269

Description: When an application uses doca_telemetry_diag to open multiple devices from a single thread, commands to the previously opened devices might fail.

Keyword: DOCA Telemetry; fwctl; multiple devices; PRM command 0x805

Detected in version: 3.2.1

5267600

Description: On BlueField-3 devices running the Oracle Linux 9 BFB, a VLAN tag configured on an OVS bridge port might be removed after a BlueField software reset (SW_RESET). Traffic that depends on the tag fails.

Keyword: OVS; VLAN tag; software reset; Oracle Linux 9

Detected in version: 3.6.0

5262094

Description: DO-SPCX might configure congestion-control (PPCC) parameters that the current deployment guide no longer specifies, which can reduce performance in some NCCL tests.

Keyword: DMS; DO-SPCX; congestion control; PPCC; performance

Detected in version: 3.5.0

5249315

Description: When the MTU of a physical function (PF) differs from the MTU of its virtual functions (VFs), OVS might block its main thread.

Keyword: Open vSwitch; MTU; PF; VF

Detected in version: 3.4.0

5246257

Description: With doca_spcx_cc, the NEW_FLOW PPCC counter of a congestion-control algorithm slot that carries no new flows might show negative values (for example, 0xfffffffd) while traffic runs on another slot.

Keyword: DOCA PCC; doca_spcx_cc; PPCC counters; NEW_FLOW

Detected in version: 3.5.0

5240356

Description: DO-SPCX profiles might apply outdated congestion-control parameter values, such as AI, HAI, and HAI_PERIOD_NS, instead of the values for the MoE congestion-control algorithm.

Keyword: DMS; DO-SPCX; congestion control; profiles

Detected in version: 3.5.0

5239300

Description: In bare-metal deployments, DO-SPCX might write /etc/modprobe.d/spcx-ib-core.conf with options ib_core netns_mode=0, a setting that bare-metal deployments don't require.

Keyword: DMS; DO-SPCX; bare metal; ib_core

Detected in version: 3.5.0

5237182

Description: After a DO-SPCX configuration phase fails, cleanup might not clear the failed state, so the next installation fails at the same phase.

Keyword: DMS; DO-SPCX; cleanup; reinstall

Detected in version: 3.5.0

5236916

Description: After a successful DO-SPCX configuration stage, rebooting the host might leave it partially configured because the plan target does not activate on boot.

Keyword: DMS; DO-SPCX; configuration; reboot

Detected in version: 3.5.0

5231406

Description: When running ib_write_bw over MRC with --tclass 0, data packets might be sent with DSCP 1 instead of the requested DSCP 0.

Keyword: MRC; perftest; DSCP; traffic class

Detected in version: 3.5.0

5231165

Description: When doca-installer terminates because no target device matches (for example, because of an OPN mismatch, a missing RShim device, or all devices being excluded), it might run a full mstdump on every BlueField device on the host.

Keyword: doca-installer; mstdump; error handling

Detected in version: 3.5.0

5230106

Description: On hosts with multiple network adapters and interfaces, Network Diag measurements might fail with timeouts because concurrent mlxlink commands contend for the same device register.

Keyword: Network Diag; scale; mlxlink; timeout

Detected in version: 3.4.0

5226978

Description: On PXE-booted hosts, such as hosts managed by Bright Cluster Manager (BCM), the DO-SPCX preparation stage might fail after the required reboot because the boot-policy check does not find the required kernel boot parameters.

Keyword: DMS; DO-SPCX; BCM; PXE boot

Detected in version: 3.5.0

5221629

Description: In the Path Selector Switching (PSS) service, a percentage_per_filter error-injection filter configured with a large path selector (PS) value might not drop packets.

Keyword: PSS; error injection; path selector; packet drop

Detected in version: 3.5.0

5221627

Description: When cyclic capture is enabled in the PSS service, capture files might exceed the configured maximum file size.

Keyword: PSS; dynamic debug; packet capture; file size

Detected in version: 3.5.0

5221386

Description: The DO-SPCX preparation stage might fail on BlueField-4 NIC devices because the profile accepts only ConnectX-8 devices.

Keyword: DMS; DO-SPCX; BlueField-4 NIC; preparation

Detected in version: 3.5.0

5220336

Description: On GB300 systems, the DO-SPCX Spectrum-XGS configuration might stop after the first PF, leaving the other PFs with their default congestion-control settings.

Keyword: DMS; DO-SPCX; Spectrum-XGS; GB300; congestion control

Detected in version: 3.5.0

5219542

Description: DOCA Lifeguard Client SDK requests might block indefinitely when the DOCA Lifeguard Service does not reply, because the SDK does not provide a request timeout or cancellation.

Keyword: DOCA Lifeguard; client SDK; timeout; cancellation

Detected in version: 3.6.0

5219541

Description: The DOCA Lifeguard Service might accept a DEPOSIT request that does not include a file descriptor and report success. The client then treats the device as protected, although the service cannot preserve the datapath if the application exits or crashes.

Keyword: DOCA Lifeguard; DEPOSIT; file descriptor

Detected in version: 3.6.0

5189291

Description: The DOCA Flow PSP master key application might access invalid stack memory when the same master key is updated twice.

Keyword: DOCA Flow; PSP; master key

Detected in version: 3.5.0

5179282

Description: DMS cleanup might report success but leave files that DO-SPCX created on the host, including /etc/modprobe.d/spcx-ib-core.conf, which keeps changing the RDMA network namespace behavior after reboot.

Keyword: DMS; DO-SPCX; cleanup; ib_core

Detected in version: 3.5.0

5171226

Description: When cleanup of a failed DMS plan fails, dms-cli might return exit code 0 although its output reports success: false.

Keyword: DMS; dms-cli; cleanup; exit code

Detected in version: 3.5.0

5170001

Description: When DOCA XPlane receives an LLDP update that contains a malformed TLV, it might treat the update as a flush and clear all remote failures learned on that port until the next periodic LLDP update.

Keyword: DOCA XPlane; LLDP; TLV; remote failures

Detected in version: 3.4.0

5156303

Description: Running dmsc --help might display the help text of a different command-line tool instead of dmsc.

Keyword: DMS; dmsc; help text

Detected in version: 3.5.0

5153283

Description: The DOCA Secure Channel application might leak memory when a consumer or producer fails to start.

Keyword: DOCA Secure Channel; DOCA Comch; memory leak

Detected in version: 3.5.0

5138631

5224972

5237084

Description: In a DOCA Flow LPM pipe, an exact-match field narrower than the word that holds it, such as the 24-bit VXLAN VNI or the 20-bit IPv6 flow label, might be left out of the match key unless the whole word is masked. Packets that should hit the entry go to the pipe's miss path, and no error is reported. In addition, a custom-header (flex parser) match field next to another masked field might also match on that field's mask bits.

Keyword: DOCA Flow; LPM pipe; exact match; flex parser

Detected in version: 3.5.0

5131124

Description: DOCA Bench might report errors while tearing down a DOCA GPUNetIO bidirectional Ethernet benchmark. The benchmark itself completes successfully.

Keyword: DOCA Bench; DOCA GPUNetIO; teardown

Detected in version: 3.5.0

4884489

4985740

4988459

Description: The DMS YANG model, and the output of dmsc generate path, might include leaves that DMS get responses do not return, such as id, name, and rp_enabled.

Keyword: DMS; YANG model; generate path

Detected in version: 3.5.0

4764229

Description: Updating the actions of a DOCA Flow ACL pipe entry might fail.

Keyword: DOCA Flow; ACL pipe; entry update

Detected in version: 3.5.0

DOCA-Host and DOCA Drivers Bug Fixes

Internal Ref.

Issue

5236231

Description: Resolved an issue where running apt full-upgrade on DGX OS 8.0.0 incorrectly pulled in GKE-flavor kernel headers (linux-headers-7.0.0-1005-gke), causing the mlnx-nvme DKMS build to fail against the GKE kernel tree and leaving the package unconfigured.

Keywords: apt full-upgrade, GKE kernel headers, mlnx-nvme, DKMS, DGX OS 8.0.0

Detected in version: 3.4.0

Fixed in release: 3.6.0

5278305 / 5298865 / NVbug 6787749

Description: Resolved an issue where, on a host with both RoCE and native InfiniBand devices, libibumad attempted to open Ethernet ports as InfiniBand devices when discovering SMI/GSI pairs.

Keywords: libibumad SMI, GSI ethernet

Detected in version: 3.2.0

Fixed in release: 3.6.0

5273313 / 5273320 / 5271745 / NVbug 6779071

Description: Resolved an issue where IPv6 route lookups in the RDMA stack ignored the bound network interface when a source address was also specified. Instead, the most specific route was selected, even if that route used a different interface. The issue affected both outgoing address/route resolution (rdma_resolve_addr) and validation of incoming RDMA CM connection requests.

Keywords: IPv6, route lookup, RDMA CM, rdma_resolve_addr, bound interface

Detected in version: 3.5.0

Fixed in release: 3.6.0

5218649 / 5218479 / NVbug 6598784

Description: Resolved an issue where the driver did not report Fractional Frequency Offset (FFO) values through the appropriate kernel API.

Keywords: FFO, DPLL, kernel API, driver

Detected in version: 3.5.0

Fixed in release: 3.6.0

5317216

Description: Resolved an issue where kernel packages built manually from source did not include the kernel version in the built package name. The fix applies to building the packages manually and not to using doca-kernel-support; the behavior of doca-kernel-support is unchanged.

Keywords: kernel packages, RPM, package name, kernel version, doca-kernel-support

Detected in version: 3.5.0

Fixed in release: 3.6.0

BSP Bug Fixes

Ref #

Details

5358445

Description: When the bf.cfg file has certain sizes, such as 8,197 bytes, BFB installation might receive a truncated copy of it, discard it, and continue the installation without the bf.cfg configuration, without reporting an error.

Keywords: bf.cfg; BFB installation; RShim

Detected in version: 4.16.0

5307749

5303810

Description: After a full BFB installation, firmware versions might not be readable from the host because the ATF mailbox service stops responding after its interrupt is disabled. As a result, doca-installer might report a successful installation as failed, and doca-installer --compare might show N/A for the ATF, UEFI, BMC, and CEC versions.

Keywords: doca-installer; ATF mailbox; firmware version query

Detected in version: 4.16.0

5273309

Description: On BlueField devices installed from the RHEL 10.2 BFB image, dnf operations might fail or time out because the image includes a repository configuration file (/etc/yum.repos.d/rh.repo) that references hosts reachable only from inside the NVIDIA network.

Keywords: RHEL; dnf; package repository; BFB

Detected in version: 4.16.0

5221877

Description: On BlueField images where OP-TEE is not enabled, the tee-supplicant service might still start at boot and fail, adding errors to the system logs.

Keywords: OP-TEE; tee-supplicant; system logs

Detected in version: 4.16.0


BMC Bug Fixes

Ref #

Issue Details

5242838

Description: On NVIDIA® BlueField®-3 and BlueField-4, downloading attachments of existing BMC dump entries through Redfish might cause D-Bus timeouts and bmcweb crashes, making the Redfish service temporarily unavailable.

Reported in version: 26.07

5218062

Description: In rare cases, a Redfish GET request to /redfish/v1/Systems/Bluefield/Oem/Nvidia might return a truncated BaseMAC value.

Reported in version: 26.07

5333640

Description: When a BFB transfer through Redfish SimpleUpdate over SCP fails because the BMC holds an earlier public key for the remote server, the BMC might report only Download failed, without indicating the cause.

Reported in version: 26.07

5322083

Description: On BlueField-2 and BlueField-3, the Redfish firmware inventory entry /redfish/v1/UpdateService/FirmwareInventory/DPU_OFED might report "Version": "N/A" instead of the BlueField OFED version.

Reported in version: TBD

BlueField-3 Firmware Bug Fixes

Internal ref.

Issue

5343627 / 5343844 / 5343877 / 5105066 / 5105068

Description: Resolved an issue where the link failed to come up with a specific LPO (linear pluggable optics) module. Dedicated SI parameters were added for this module.

Keywords: LPO, SI parameters, link

Detected in version:

32.50.1002

Fixed in release:

32.51.1002

5093033 / 5320893 / 5275267 / 5299597

Description: Resolved an issue where the BlueField-3 PCIe link could go down unexpectedly during the Itloop Sequential Speed Change test.

Keywords: BlueField-3, PCIe link, Itloop Sequential Speed Change test

Detected in version:

32.43.4100

Fixed in release:

32.51.1002

5350445

Description: Resolved an issue where an active congestion control flow context might be outdated unexpectedly when a large number of congestion control flows are active.

Keywords: PCC

Detected in version:

32.50.1002

Fixed in release:

32.51.1002

5292838 / 5259132

Description: Resolved a fatal firmware assert (FATAL_ERR_FW_INTERNAL_ERR, ext_synd 0x8257) that occurred in hostless DPA configurations on BlueField-3. NCORE Firewall Profiles were added to cover Pseudo and Page Supplier Function GVMIs.

Keywords: FATAL_ERR_FW_INTERNAL_ERR, DPA, NCORE, Firewall Profiles, GVMI, UTOPX

Detected in version:

32.50.1002

Fixed in release:

32.51.1002

5257848 / 5318876 / NVBug 6720743

Description: Resolved an issue where querying congestion counters on a DPU host returned the congestion counters of the external host instead.

Keywords: Congestion counters, DPU host, external host

Detected in version:

32.47.1026

Fixed in release:

32.51.1002

5235696

Description: Resolved incorrect handling of MACsec confidentiality-offset values during conversion between the software interface and the hardware SADB, which could result in improper packet encryption.

Keywords: MACsec, confidentiality-offset, SADB, packet encryption

Detected in version:

32.50.1002

Fixed in release:

32.51.1002

5224818 / 5231776 / 4403151

Description: Resolved an issue where RDMA READ bandwidth could drop sharply and fail to recover at high QP counts when DCQCN and Adaptive Retransmission were both enabled. Duplicate-READ NACK events are no longer treated by DCQCN as congestion signals.

Keywords: RDMA READ, BW drop, DCQCN, Adaptive Retransmission, QP

Detected in version:

32.50.1002

Fixed in release:

32.51.1002

5170451 / 5170462

Description: Resolved an issue where, with PCIe switch emulation enabled on BlueField-3, the host could hang during boot or fail to load the mlx5 driver after a cold reboot on systems with IOMMU enabled.

Keywords: VirtIO, mlxconfig, cold reboot, Dell PowerEdge, PF load

Detected in version:

32.50.1002

Fixed in release:

32.51.1002

5157189 / 4866555

Description: Resolved an issue where, when massive flows occurred in parallel with PCC, INIT2RTR could report a false alarm with syndrome 0x7 "dead IRISC."

Keywords: IRISC, dead IRISC, PCC, INIT2RTR, syndrome 0x7

Detected in version:

32.47.1026

Fixed in release:

32.51.1002

5125165

Description: Resolved an incorrect default value for the "Auto Slot Power Limit Disable" bit in the PCIe Slot Control Register. The bit was erroneously set to 1, preventing the following BlueField-3 controllers SKUs (900-9D3C6-00SV-DA and 900-9D3C6-00CV-DA0) from generating downstream Set Slot Power Limit messages.

Keywords: PCIe Slot Control, power limit, BlueField-3, downstream

Detected in version:

32.49.1014

Fixed in release:

32.51.1002

5093878 / 5138237

Description: Resolved a credit leak and resulting TX timeouts that occurred when a DPA application used cross-VHCA, cross-host doorbell messaging via DPA Outbox.

Keywords: DPA Outbox, credit leak, TX timeout, cross-VHCA, doorbell

Detected in version:

32.46.1006

Fixed in release:

32.51.1002

5081175

Description: Resolved a ZTR RTT CC issue in topology-aware mode where fabric round-trip times exceeding 200 µs caused false congestion signals and sustained below-line-rate throughput.

Keywords: ZTR RTT CC, topology-aware, fabric RTT, congestion

Detected in version:

32.50.1002

Fixed in release:

32.51.1002

5079086 / 5096997

Description: Resolved an issue where emulated NVMe hot-plug and hot-unplug operations on BlueField-3 did not complete while the host was powered off.

Keywords: NVMe hotplug, plug, unplug, host power-off

Detected in version:

32.49.1014

Fixed in release:

32.51.1002

5035649

Description: When Bus Master Enable (BME) is cleared for a virtio-blk device that uses transaction layer packet (TLP) emulation, host I/O can continue instead of stopping.

Keywords: BME, virtio-blk, TLP emulation, host I/O

Detected in version:

32.49.1014

Fixed in release:

32.51.1002

5019624

Description: Resolved an issue where a PCIe SBR did not reset the MCTP discovery state, causing the device to miss EndpointDiscovery requests.

Keywords: PCIe SBR, MCTP, EndpointDiscovery, reset

Detected in version:

32.50.1002

Fixed in release:

32.51.1002

4728626 / 4939133

Description: Occasionally, after rebooting the DPU, one of the ports may become inaccessible.

Keywords: DPU reboot, port inaccessible, VSC, restricted mode

Detected in version:

32.47.1026

Fixed in release:

32.51.1002

BlueField-2 Firmware Bug Fixes

Internal ref.

Issue

5305143

Description: Resolved an issue where RDE reads of next-boot settings during firmware burn or reactivation caused a semi-fatal assert (0x83df) when the flash was busy. The firmware now returns NOT_READY to the BMC, which can retry the request after the flash operation completes.

Keywords: RDE, flash, firmware reactivation, semi-fatal assert, 0x83df, NOT_READY, BMC, PLDM

Detected in version:

24.50.1002

Fixed in release:

24.51.1002

5157189 / 4866555

Description: Resolved an issue where, when massive flows occurred in parallel with PCC, INIT2RTR could report a false alarm with syndrome 0x7 "dead IRISC."

Keywords: IRISC, dead IRISC, PCC, INIT2RTR, syndrome 0x7

Detected in version:

24.47.1026

Fixed in release:

24.51.1002

5019624

Description: Resolved an issue where a PCIe SBR did not reset the MCTP discovery state, causing the device to miss EndpointDiscovery requests.

Keywords: PCIe SBR, MCTP, EndpointDiscovery, reset

Detected in version:

24.50.1002

Fixed in release:

24.51.1002

4728626 / 4939133

Description: Occasionally, after rebooting the DPU, one of the ports may become inaccessible.

Keywords: DPU reboot, port inaccessible, VSC, restricted mode

Detected in version:

24.47.1026

Fixed in release:

24.51.1002

Last updated: