Intel® Ethernet Controller Products 27.3 Release Notes

2.1.1 General

2.1.2 Linux Driver

• Prior to irdma version 1.8.45, installing the OOT irdma driver on a system with RDMA-capable Intel® Ethernet Connection X722/Intel® Ethernet Network Adapter X722 ports and using an OS or kernel with an in-tree irdma driver could cause a system crash. To prevent a system crash when using OOT irdma drivers, either use irdma 1.8.45, or update i40e version (2.18.9 or greater) and load it before this new irdma is loaded.
• AF_XDP based applications may cause system crash on packet receive with RHEL based 4.18 kernels.
• During a long reboot cycle test (about 250-500 reboots) of the Intel Ethernet 800 Series adapters, the Intel ICE and iavf driver may experience kernel panics leading to an abnormal reboot of the server.
• The commands ethtool -C [rx|tx]-frames are not supported by the iavf driver and will be ignored.
• Setting [tx|rx]-frames-irq using ethtool -C may not correctly save the intended setting and may reset the value back to the default value of 0.
• Interrupt Moderation settings reset to default when the queue settings of a port are modified using the ethtool -L ethx combined XX command.
• When a VM is running heavy traffic loads and is attached to a Virtual Switch with either SR-IOV enabled or VMQ offload enabled, repeatedly enabling and disabling the SR-IOV/VMQ setting on the vNIC in the VM may result in a BSOD.

Linux RDMA Driver

• In order to send or receive RDMA traffic, the network interface associated with the RDMA device must be up. If the network interface experiences a link down event (for example, a disconnected cable or ip link set <interface> down), the associated RDMA device is removed and no longer available to RDMA applications. When the network interface link is restored, the RDMA device is automatically re-added.
• RHEL 8.5 only: Any usermode test that uses ibv_create_ah (For example, a RoCEv2 usermode test such as udaddy) will fail.
• Due to a nondeterministic race condition, if the irdma driver is loaded in Linux by an Intel® Ethernet 800 series device with non-standard MTU (i.e., non-1500B MTU), the system's network interfaces may fail to load after reboot. After failing to load, interactions with the networking stack may hang on the system. Multiple reboots may be required to avoid the condition.
• The Devlink command devlink dev param show (DEVLINK_CMD_PARAM_GET) does not report MinSREV values for firmware (fw.mgmt.srev) and OROM (fw.undi.srev). This defect was also seen on the NVMUpdate tool, which caused an inventory error.

2.1.3 Windows Driver

• When a VM is running heavy traffic loads and is attached to a Virtual Switch with either SR-IOV enabled or VMQ offload enabled, repeatedly enabling and disabling the SR-IOV/VMQ setting on the vNIC in the VM, may result a VM freeze/hang.

3.1.1 General

• Properties that can be modified through the manageability sideband interface PLDM Type 6: RDE, such as EthernetInterface->AutoNeg or NetworkPort->FlowControlConfiguration do not possess a permanent storage location on internal memory. Changes made through RDE are not preserved following a power cycle/PCI reset.
• Link issues (for example, false link, long time-to-link (TTL), excessive link flaps, no link) may occur when the Parkvale (C827/XL827) retimer is interfaced with SX/LX, SR/LR, SR4/LR4, AOC limiting optics. This issue is isolated to Parkvale line side PMD RX susceptibility to noise.
• Intel Ethernet 800 Series adapters in 4x25GbE or 8x10GbE configurations will be limited to a maximum total transmit bandwidth of roughly 28Gbps per port for 25GbE ports and 12Gbps per port on 10GbE ports.
• This maximum is a total combination of any mix of network (leaving the port) and loopback (VF -> VF/VF -> PF/PF ->VF) TX traffic on a given port and is designed to allow each port to maintain port speed transmit bandwidth at the specific port speed when in 25GbE or 10GbE mode.
• If the PF is transmitting traffic as well as the VF(s), under contention the PF has access to up to 50% TX bandwidth for the port and all VFs have access to 50% bandwidth for the port, which will also impact the total available bandwidth for forwarding.
• Note: When calculating the maximum bandwidth under contention for bi-directional loopback traffic, the number of TX loopback actions are twice that of a similar unidirectional loopback case, since both sides are transmitting.
• The version of the Ethernet Port Configuration Tool available in Release 26.1 may not be working as expected. This has been resolved in Release 26.4.
• E810 currently supports a subset of 1000BASE-T SFP module types, which use SGMII to connect back to the E810. In order for the E810 to properly know the link status of the module's BASE-T external connection, the module must indicate the BASE-T side link status to the E810. An SGMII link between E810 and the 1000BASE-T SFP module allows the module to indicate its link status to the E810 using SGMII Auto Negotiation. However 1000BASE-T SFP modules implement this in a wide variety of ways, and other methods which do not use SGMII are currently unsupported in E810. Depending on the implementation, link may never be achieved. In other cases, if the module sends IDLEs to the E810 when there is no BASE-T link, the E810 may interpret this as a link partner sending valid data and may show link as being up even though it is only connected to the module and there is no link on the module's BASE-T external connection.
• If the PF has no link then a Linux VM previously using a VF will not be able to pass traffic to other VMs without the patch found here. Link to patch This patch routes packets to the virtual interface.
• Note: This is a permanent 3rd party issue. No expected action on the part of Intel.
• Some devices support auto-negotiation. Selecting this causes the device to advertise the value stored in its NVM (usually disabled).
• VXLAN switch creation on Windows Server 2019 Hyper V might fail.

3.1.2 Firmware

• Promiscuous mode does not see all packets: it sees only those packets arriving over the wire (that is, not sent from the same physical function (PF) but a different virtual function (VF).
• Per the specification, the Get LLDP command (0x28) response may contain only 2 TLVs (instead of 3).
• When software is requesting from firmware the port parameters on port 0 (via AQ the connectivity type), the response is BACKPLANE_CONNECTIVITY, when it should be CAGE_CONNECTIVITY
• Health status messages are not cleared with a PF reset, even after the reported issue is resolved.
• Flow control settings have no effect on traffic, and counters do not increment with flow control set to TX=ON and Rx=OFF. However, flow control works fine with values set to TX=On RX=ON.

3.1.3 Linux Driver

• Linux sysctl commands, or any automated scripting that alerts or sets /proc/sys/ attributes using sysctl, might encounter a system crash that includes irdma_net_event in the dmesg stack trace.
• Workaround: With OOT irdma-1.8.X installed on the system, avoid running sysctl while drivers are being loaded or unloaded.
• VXLAN stateless offloads (checksum, TSO), as well as TC filters directing traffic to a VXLAN interface are not supported with Linux v5.9 or later.
• Linux ice driver 1.2.1 cannot be compiled with E810 3.2 NVM images. The version on the kernel is 5.15.2.
• On RHEL8.5, l2-fwd-offload cannot be turned on.
• When spoofchk is turn on, the VF device driver will have pending DMA allocations while it is released from the device.
• After changing link speed to 1G on the E810-XXVDA4, the PF driver cannot detect a link up on the adapter. As a workaround the user can force 1G on the second side.
• If the rpmbuild command of the new iavf version fails due to the existing auxiliary files installed, please use --define "_unpackaged_files_terminate_build 0" with the rpmbuild command. Usage/Workaround will look like rpmbuild -tb iavf-4.4.0_rc53.tar.gz --define "_unpackaged_files_terminate_build 0" ".
• irdma stops working if the number of ice driver queues are changed (ethtool -L) while the irdma driver is loaded. As a workaround, remove (if previously loaded) and reload irdma after changing the number of queues.
• When the queue settings of a port are modified using the ethtool -L ethx combined XX command, the Interrupt Moderation settings reset to default.
• When using bonding mode 5 (i.e., balance-tlb or adaptive transmit load balancing), if you add multiple VFs to the bond, they are assigned duplicate MAC address. When the VFs are joined with the bond interface, the Linux bonding driver sets the MAC address for the VFs to the same value. The MAC address is based on the first active VF added to that bond. This results in balance-tlb mode not functioning as expected. PF interfaces behave as expected.
• The presence of duplicate MAC addresses may cause further issues, depending on your switch configuration.
• When the maximum allowed number of VLAN filters are created on a trusted VF, and the VF is then set to untrusted and the VM is rebooted, the iavf driver may not load correctly in the VM and may show errors in the VM dmesg log.
• Changing the FEC value from BaseR to RS results in an error message in dmesg, and may result in link issues.
• UEFI PXE installation of Red Hat Enterprise Linux 8.4 on a local disk results with the system failing to boot.
• When a VF interface is set as 'up' and assigned to a namespace, and the namespace is then deleted, the dmesg log may show the error Failed to set LAN Tx queue context, error: ICE_ERR_PARAM followed by error codes from the ice and iavf drivers.