Description

The software is the driver of Mellanox NIC in the Linux operating system (OS).

Installation steps

Restrictions and guidelines

1.Before updating the software, as a best practice, understand the changes on the software between versions, assess the impacts of the changes on the services and read supporting materials.

2.In systems with RHEL 8.4 and above kernel versions, installing Mellanox-OFED 5.4-0.5.1.1 or higher version driver package will cause abnormal CIFS module, resulting in some functions unable to be used normally. For example, when using the CIFS module to mount a disk, it will prompt "unsupported file system".

3.In the H3Linux 2.0.2-SP01 and NingOS V3 1.0.2306, please use " ./mlnxofedinstall --without-fw-update --force --distro openeuler20.03sp3 " for driver installation and update.

Installation steps

OFED package

1.Log in to the server system by using the administrator account and upload the driver package file.tar.gz to the system.

2.Decompress the driver source code package by executing the following command: tar -zxvf file.tar.gz.

(Note: The common commands for decompressing a compressed package are as follows:

tar –xvf file.tar

tar –zxvf file.tgz

tar –jxvf file.bz2

unrar e file.rar

unzip file.zip)

3.Execute the command chmod –R 777 * to grant execute permission for decompressed files.

4.Execute the command ./mlnxofedinstall --without-fw-update --force to upgrade driver.

The following prompts may appear when the driver is installed

This may be caused by a system kernel version upgrade. Please add the --add-kernel-support parameter to compile and install the driver based on the current kernel version.

Publish Description

Supported products

H3C UniServer B5700 G3

H3C UniServer B5700 G5

H3C UniServer B5800 G3

H3C UniServer B7800 G3

H3C UniServer E3200 G3

H3C UniServer R2700 G3

H3C UniServer R2900 G3

H3C UniServer R4300 G3

H3C UniServer R4300 G5

H3C UniServer R4330 G5

H3C UniServer R4330 G5 H3

H3C UniServer R4500 G3

H3C UniServer R4700 G3

H3C UniServer R4700 G5

H3C UniServer R4700 G6

H3C UniServer R4900 G3

H3C UniServer R4900 G5

H3C UniServer R4900 G6

H3C UniServer R4900 G6 Ultra

H3C UniServer R4930 G5

H3C UniServer R4930 G5 H3

H3C UniServer R4950 G5

H3C UniServer R4950 G5 Milan

H3C UniServer R4950 G5 Rome

H3C UniServer R4950 G6

H3C UniServer R5300 G3

H3C UniServer R5300 G5

H3C UniServer R5500 G5 AMD

H3C UniServer R5500 G5 Intel

H3C UniServer R6700 G3

H3C UniServer R6900 G3

H3C UniServer R8900 G3

H3C UniServer R5350 G6

H3C UniServer R6900 G5

H3C UniServer R5500 G6 AMD

H3C UniServer R5500 G6 Intel

H3C UniServer R5300 G6

H3C UniServer R5500LC G5 Intel

H3C UniServer R5500 G5 HYGON

H3C UniServer B5700 G6

H3C UniServer R4900LE G6 Ultra

H3C UniServer R4700LE G6

H3C UniServer R4900LC G5

H3C UniServer R4700LC G5

H3C UniServer R4930LC G5 H3

H3C UniServer R4300 G6

H3C UniServer R4500 G6

H3C UniServer R6700 G6

H3C UniServer R6900 G6

H3C UniServer R5330 G7

H3C UniServer R5500 G7 Hygon

H3C UniServer R4930 G7

H3C UniServer E3300 G6

H3C UniServer R3950 G6

H3C UniServer R4330 G7

H3C UniServer R4950 G7

H3C UniServer R3950 G7

H3C UniServer R3350 G7

H3C UniServer R4900 G7

H3C UniServer R4700 G7

H3C UniServer R3900 G7

H3C UniServer R4300 G7

Restrictions and guidelines

1.Before updating the software, as a best practice, understand the changes on the software between versions, assess the impacts of the changes on the services and read supporting materials.

2.In systems with RHEL 8.4 and above kernel versions, installing Mellanox-OFED 5.4-0.5.1.1 or higher version driver package will cause abnormal CIFS module, resulting in some functions unable to be used normally. For example, when using the CIFS module to mount a disk, it will prompt "unsupported file system".

3.In the H3Linux 2.0.2-SP01 and NingOS V3 1.0.2306, please use " ./mlnxofedinstall --without-fw-update --force --distro openeuler20.03sp3 " for driver installation and update.

Problems Solved

1.Fixed the problem that the driver did not consider whether the device was in the disassembly or detection process when performing recovery operations, which would cause a kernel crash

2.Fixed the problem that setting the rdma_cm RoCE static rate to 0 would lead to high storage IO latency when a large number of rdma_cm connections were established

3.Fixed the issue that the "rdma res show qp" command would return an "Invalid argument" error when there were a large number of QPs

4.Upgrade knem to support RHEL 8.7 system, so as to avoid the problem of call tracking caused by soft lock

5.Fixed an issue where the system might hang during startup after loading the mlx5_core driver in some systems with multiple ConnectX adapters

6.Added support for using VF directly and VF assigned to QEMU to use Relaxed Ordering. Relaxed Ordering can improve performance under certain settings. But so far, it's only available in PF.

New Functions

1.Added support for RHEL 9.2, RHEL 8.8, SLES15-SP5, KYLIN10-SP3 networks.

Version:24.10-1.1.4.0(2026/4/14)

Upgrade recommendations

Upgrade the driver,as a best practice.

Problems Solved

New Functions

Version:25.10-1.2.2.0(2026/1/30)

Upgrade recommendations

Upgrade the driver,as a best practice.

Problems Solved

New Functions

Version:24.10-2.1.8.0(2025/4/7)

Upgrade recommendations

Upgrade the driver,as a best practice.

Problems Solved

New Functions

Version:23.10-3.2.2.0(2024/9/24)

Upgrade recommendations

Upgrade the driver,as a best practice.

Problems Solved

1.Fixed the setting of ATS for DMABUF MRs that caused some MRs to miss the ATS enablement. Lack of ATS enablement on DMABUF MRs results in slower performance when using these MRs.

2.On rare occasions, rdmacm applications could not find the device upon creating new RDMA devices, as the CMA driver lost some of the devices due to an overflow issue.

3.Fixed an issue that could have caused memory corruption when running XDP traffic.

4.Fixed an issue that prevented the tc_wrap tool from properly working when VLAN is configured as the tool wrongly handled the library function return value.

5.Fixed an issue that caused a deadlock in the flow of disabling the LAG when changing eswitch mode from switchdev to legacy when a LAG bond existed on the machine.

6.Fixed an issue related to the driver's internal MR cache cleanup that caused high memory consumption on the host.

New Functions

1.Added support for the following OSes:RHEL 8.10、RHEL 9.4、Debian 12.5

Version:23.10-1.1.9.0(2024/1/5)

Upgrade recommendations

Upgrade the driver,as a best practice.

Problems Solved

1.Resolved an issue on ConnectX-4 Lx, where the VF state was not configured correctly following the activation of SR-IOV.

New Functions

1.None.

Version:5.8-3.0.7.0(2023/10/16)

Upgrade recommendations

Upgrade the driver,as a best practice.

Problems Solved

1.Fixed the problem that the driver did not consider whether the device was in the disassembly or detection process when performing recovery operations, which would cause a kernel crash

2.Fixed the problem that setting the rdma_cm RoCE static rate to 0 would lead to high storage IO latency when a large number of rdma_cm connections were established

3.Fixed the issue that the "rdma res show qp" command would return an "Invalid argument" error when there were a large number of QPs

4.Upgrade knem to support RHEL 8.7 system, so as to avoid the problem of call tracking caused by soft lock

5.Fixed an issue where the system might hang during startup after loading the mlx5_core driver in some systems with multiple ConnectX adapters

6.Added support for using VF directly and VF assigned to QEMU to use Relaxed Ordering. Relaxed Ordering can improve performance under certain settings. But so far, it's only available in PF.

New Functions

1.Added support for RHEL 9.2, RHEL 8.8, SLES15-SP5, KYLIN10-SP3 networks.

Version:5.8-1.1.2.1(2023/8/21)

Upgrade recommendations

Upgrade the driver,as a best practice.

Problems Solved

1.The knem-dkms package explicitly requires GCC to build the knem driver (at install times). Under some circumstances, on Debian systems, the apt install method may result in a system that has only gcc-<version> (e.g., gcc-10) installed.

2.The host driver probe did not check whether there are existing SFs which are present in the device, causing the host driver to not recreate those SFs.

3.Installing MLNX_OFED_LINUX on an Ubuntu system with CUDA (version < 11.6) may result in an automatic installation of the ucx-cuda package that will fail with an error message in the log file ucx-cuda.debinstall.log about missing dependencies.

4. When using BlueField with old kernels, multiple OVS meter do not work.

5.Upgrading from version 5.6-2 to 5.7 failed.

6.Starting firmware version xx.34.0350, a new NVCONFIG has been added to the ARM side only: MANAGEMENT_PF_MODE.If this config is on, the user will see a PCI Function (PF) which failed to probe:

[ 6.837102] mlx5_core 0000:03:00.2: mlx5_cmd_check:756:(pid 206): ENABLE_HCA(0x104) op_mod(0x0) failed, status bad parameter(0x3), syndrome (0x6ca1f5)

[ 6.864227] mlx5_core 0000:03:00.2: mlx5_peer_pf_init:40:(pid 206): Failed to enable peer PF HCA err(-22)

[ 6.883453] mlx5_core 0000:03:00.2: mlx5_load:1129:(pid 206): Failed to init embedded CPU

[ 8.261268] mlx5_core 0000:03:00.2: init_one:1365:(pid 206): mlx5_load_one failed with error code -22

[ 8.280056] mlx5_core: probe of 0000:03:00.2 failed with error -22

7.Upgrading from version 5.6-2 to 5.7 failed.

8.When using sysfs to read the hash function used to distribute the traffic between the TIRs (Transport Interface Receive), on occasion, the server crashed.

9.STP packets failed to be transmitted.

10.HW-GRO feature was blocked due to firmware limitations.

11.In older kernel version, PTP was not supported over VLAN interfaces.

12.The rdma res show cm_id command does not list all cm_ids when some of them are in LISTEN state.

13.Creating and deleting MRs, caused a kernel slab cache leak issue.

New Functions

1.[All HCAs] Removed dependency between SR-IOV and eSwitch mode. Currently, there are three eSwitch modes: none, legacy, and switchdev (non of which are the default mode). When disabling SR-IOV, the current eSwitch mode will be changed to none. This feature removes eSwitch mode none and also removes dependency between SR-IOV and eSwitch mode.

2.[All HCAs] Added support for running DevLink commands in parallel on different DevLink devices is possible. For example, burning firmware on a few cards on the same host in parallel using DevLink API is now possible.