概述
========
此驱动程序支持内核版本3.10.0及更新版本。但是，有些功能可能需要更新的内核版本。关联的虚拟功能（VF）驱动程序对于这个驱动程序，是iavf。可以使用ethtool、lspci和ip获取驱动程序信息。
更新ethtool时，说明书可在在本文件中稍后的“附加配置”部分中找到。
目前仅支持将此驱动程序作为可加载模块。英特尔不是针对内核源提供补丁。
有关硬件要求的问题，请参阅随英特尔适配器提供的文档。列出的所有硬件要求均适用于Linux。
此驱动程序支持内核4.14及更高版本的 XDP (Express Data Path)和内核4.18及更高版本上的AF_XDP zero-copy零拷贝。请注意，XDP被阻止用于大于3KB的大小的帧。
识别适配器
========================
驱动程序与以下设备兼容：
*英特尔（R）以太网控制器E810-C
*英特尔（R）以太网控制器E810-XXV
有关如何识别适配器的信息，以及最新的英特尔网络驱动程序，请参阅英特尔技术支持网站：
http://www.intel.com/support
重要注意事项
===============
配置SR-IOV以提高网络安全性
------------------------------------------------
在虚拟化环境中，在Intel（R）以太网网络适配器上支持SR-IOV时，虚拟功能（VF）可能会受到恶意行为的影响。软件生成的层二帧，如IEEE 802.3x（链路流控制），IEEE802.1Qbb（基于优先级的流量控制）和其他此类控制不适用，预期并可以限制主机和虚拟交换机之间的通信量，降低性能。解决此问题，并确保与意外流量流，为VLAN标记配置所有启用SR-IOV的端口
从PF上的管理界面。此配置允许要删除的意外且可能是恶意的帧。
请参阅本章后面的“在启用SR-IOV的适配器端口上配置VLAN标记”
配置说明的自述文件。
如果带有活动VM的VF绑定到端口驱动程序，则不要卸载端口驱动程序
-------------------------------------------------------------
如果虚拟功能（VF）具有活动的虚拟端口，请不要卸载端口的驱动程序
计算机（VM）已绑定到它。这样做将导致端口看起来挂起。
一旦VM关闭或以其他方式释放VF，命令将完成。
固件恢复模式
----------------------
如果设备检测到以下问题，则将进入固件恢复模式：
需要重新编程固件。当设备处于固件恢复状态时
模式，它不会通过流量或允许任何配置；你只能尝试
恢复设备的固件。请参阅英特尔（R）以太网适配器和
设备用户指南，了解有关固件恢复模式和如何恢复的详细信息
从它。
Important notes for SR-IOV, RDMA, and Link Aggregation
------------------------------------------------------
Link Aggregation is mutually exclusive with SR-IOV and RDMA.
- If Link Aggregation is active, RDMA peers will not be able to register with
the PF, and SR-IOV VFs cannot be created on the PF.
- If either RDMA or SR-IOV is active, you cannot set up Link Aggregation on the
interface.
Bridging and MACVLAN are also affected by this. If you wish to use bridging or
MACVLAN with RDMA/SR-IOV, you must set up bridging or MACVLAN before enabling
RDMA or SR-IOV. If you are using bridging or MACVLAN in conjunction with SR-IOV
and/or RDMA, and you want to remove the interface from the bridge or MACVLAN,
you must follow these steps:
1. Remove RDMA if it is active
2. Destroy SR-IOV VFs if they exist
3. Remove the interface from the bridge or MACVLAN
4. Reactivate RDMA and recreate SRIOV VFs as needed
Building and Installation
=========================
The ice driver requires the Dynamic Device Personalization (DDP) package file
to enable advanced features (such as dynamic tunneling, Flow Director, RSS, and
ADQ, or others). The driver installation process installs the default DDP
package file and creates a soft link ice.pkg to the physical package
ice-x.x.x.x.pkg in the firmware root directory (typically /lib/firmware/ or
/lib/firmware/updates/). The driver install process also puts both the driver
module and the DDP file in the initramfs/initrd image.
NOTE: When the driver loads, it looks for intel/ice/ddp/ice.pkg in the firmware
root. If this file exists, the driver will download it into the device. If not,
the driver will go into Safe Mode where it will use the configuration contained
in the device's NVM. This is NOT a supported configuration and many advanced
features will not be functional. See "Dynamic Device Personalization" later for
more information.
To build a binary RPM package of this driver
--------------------------------------------
Note: RPM functionality has only been tested in Red Hat distributions.
1. Run the following command, where <x.x.x> is the version number for the
driver tar file.
# rpmbuild -tb ice-<x.x.x>.tar.gz
NOTE: For the build to work properly, the currently running kernel MUST
match the version and configuration of the installed kernel sources. If
you have just recompiled the kernel, reboot the system before building.
2. After building the RPM, the last few lines of the tool output contain the
location of the RPM file that was built. Install the RPM with one of the
following commands, where <RPM> is the location of the RPM file:
# rpm -Uvh <RPM>
or
# dnf/yum localinstall <RPM>
NOTES:
- To compile the driver on some kernel/arch combinations, you may need to
install a package with the development version of libelf (e.g. libelf-dev,
libelf-devel, elfutilsl-libelf-devel).
- When compiling an out-of-tree driver, details will vary by distribution.
However, you will usually need a kernel-devel RPM or some RPM that provides the
kernel headers at a minimum. The RPM kernel-devel will usually fill in the link
at /lib/modules/'uname -r'/build.
To manually build the driver
----------------------------
1. Move the base driver tar file to the directory of your choice.
For example, use '/home/username/ice' or '/usr/local/src/ice'.
2. Untar/unzip the archive, where <x.x.x> is the version number for the
driver tar file:
# tar zxf ice-<x.x.x>.tar.gz
3. Change to the driver src directory, where <x.x.x> is the version number
for the driver tar:
# cd ice-<x.x.x>/src/
4. Compile the driver module:
# make install
The binary will be installed as:
/lib/modules/<KERNEL VER>/updates/drivers/net/ethernet/intel/ice/ice.ko
The install location listed above is the default location. This may differ
for various Linux distributions.
NOTE: To build the driver using the schema for unified ethtool statistics
defined in https://sourceforge.net/p/e1000/wiki/Home/, use the following
command:
# make CFLAGS_EXTRA='-DUNIFIED_STATS' install
NOTE: To compile the driver with ADQ (Application Device Queues) flags set,
use the following command, where <nproc> is the number of logical cores:
# make -j<nproc> CFLAGS_EXTRA='-DADQ_PERF_COUNTERS' install
(This will also apply the above 'make install' command.)
5. Load the module using the modprobe command.
To check the version of the driver and then load it:
# modinfo ice
# modprobe ice
Alternately, make sure that any older ice drivers are removed from the
kernel before loading the new module:
# rmmod ice; modprobe ice
NOTE: To enable verbose debug messages in the kernel log, use the dynamic debug
feature (dyndbg). See "Dynamic Debug" later in this README for more information.
6. Assign an IP address to the interface by entering the following,
where <ethX> is the interface name that was shown in dmesg after modprobe:
# ip address add <IP_address>/<netmask bits> dev <ethX>
7. Verify that the interface works. Enter the following, where IP_address
is the IP address for another machine on the same subnet as the interface
that is being tested:
# ping <IP_address>
Command Line Parameters
=======================
The only command line parameter the ice driver supports is the debug parameter
that can control the default logging verbosity of the driver. (Note: dyndbg
also provides dynamic debug information.)
In general, use ethtool and other OS-specific commands to configure
user-changeable parameters after the driver is loaded.
Additional Features and Configurations
======================================
ethtool
-------
The driver utilizes the ethtool interface for driver configuration and
diagnostics, as well as displaying statistical information. The latest ethtool
version is required for this functionality. Download it at:
https://kernel.org/pub/software/network/ethtool/
NOTE: The rx_bytes value of ethtool does not match the rx_bytes value of
Netdev, due to the 4-byte CRC being stripped by the device. The difference
between the two rx_bytes values will be 4 x the number of Rx packets. For
example, if Rx packets are 10 and Netdev (software statistics) displays
rx_bytes as "X", then ethtool (hardware statistics) will display rx_bytes as
"X+40" (4 bytes CRC x 10 packets).
Viewing Link Messages
---------------------
Link messages will not be displayed to the console if the distribution is
restricting system messages. In order to see network driver link messages on
your console, set dmesg to eight by entering the following:
# dmesg -n 8
NOTE: This setting is not saved across reboots.
Dynamic Device Personalization
------------------------------
Dynamic Device Personalization (DDP) allows you to change the packet processing
pipeline of a device by applying a profile package to the device at runtime.
Profiles can be used to, for example, add support for new protocols, change
existing protocols, or change default settings. DDP profiles can also be rolled
back without rebooting the system.
The ice driver automatically installs the default DDP package file during
driver installation. NOTE: It's important to do 'make install' during initial
ice driver installation so that the driver loads the DDP package automatically.
The DDP package loads during device initialization. The driver looks for
intel/ice/ddp/ice.pkg in your firmware root (typically /lib/firmware/ or
/lib/firmware/updates/) and checks that it contains a valid DDP package file.
If the driver is unable to load the DDP package, the device will enter Safe
Mode. Safe Mode disables advanced and performance features and supports only
basic traffic and minimal functionality, such as updating the NVM or
downloading a new driver or DDP package. Safe Mode only applies to the affected
physical function and does not impact any other PFs. See the "Intel(R) Ethernet
Adapters and Devices User Guide" for more details on DDP and Safe Mode.
NOTES:
- If you encounter issues with the DDP package file, you may need to download
an updated driver or DDP package file. See the log messages for more
information.
- The ice.pkg file is a symbolic link to the default DDP package file installed
by the Linux-firmware software package or the ice out-of-tree driver
installation.
- You cannot update the DDP package if any PF drivers are already loaded. To
overwrite a package, unload all PFs and then reload the driver with the new
package.
- Only the first loaded PF per device can download a package for that device.
You can install specific DDP package files for different physical devices in
the same system. To install a specific DDP package file:
1. Download the DDP package file you want for your device.
2. Rename the file ice-xxxxxxxxxxxxxxxx.pkg, where 'xxxxxxxxxxxxxxxx' is the
unique 64-bit PCI Express device serial number (in hex) of the device you want
the package downloaded on. The filename must include the complete serial number
(including leading zeros) and be all lowercase. For example, if the 64-bit
serial number is b887a3ffffca0568, then the file name would be
ice-b887a3ffffca0568.pkg.
To find the serial number from the PCI bus address, you can use the following
command:
# lspci -vv -s af:00.0 | grep -i Serial
Capabilities: [150 v1] Device Serial Number b8-87-a3-ff-ff-ca-05-68
You can use the following command to format the serial number without the
dashes:
# lspci -vv -s af:00.0 | grep -i Serial | awk '{print $7}' | sed s/-//g
b887a3ffffca0568
3. Copy the renamed DDP package file to /lib/firmware/updates/intel/ice/ddp/.
If the directory does not yet exist, create it before copying the file.
4. Unload all of the PFs on the device.
5. Reload the driver with the new package.
NOTE: The presence of a device-specific DDP package file overrides the loading
of the default DDP package file (ice.pkg).
RDMA (Remote Direct Memory Access)
----------------------------------
Remote Direct Memory Access, or RDMA, allows a network device to transfer data
directly to and from application memory on another system, increasing
throughput and lowering latency in certain networking environments.
The ice driver supports the following RDMA protocols:
- iWARP (Internet Wide Area RDMA Protocol)
- RoCEv2 (RDMA over Converged Ethernet)
The major difference is that iWARP performs RDMA over TCP, while RoCEv2 uses
UDP.
For detailed installation and configuration information, see the README file in
the RDMA driver tarball.
Notes:
- Devices based on the Intel(R) Ethernet 800 Series do not support RDMA when
operating in multiport mode with more than 4 ports.
- You cannot use RDMA or SR-IOV when link aggregation (LAG)/bonding is active,
and vice versa. To enforce this, on kernels 4.5 and above, the driver checks
for this mutual exclusion. On kernels older than 4.5, the driver cannot check
for this exclusion and is unaware of bonding events.
NVM Express* (NVMe) over TCP and Fabrics
----------------------------------------
RDMA provides a high throughput, low latency means to directly access NVM
Express* (NVMe*) drives on a remote server.
Refer to the following for details on supported operating systems and how to
set up and configure your server and client systems:
- NVM Express over TCP for Intel(R) Ethernet Products Configuration Guide
- NVM Express over Fabrics for Intel(R) Ethernet Products with RDMA
Configuration Guide
Both guides are available on the Intel Technical Library at:
https://www.intel.com/content/www/us/en/design/products-and-solutions/networking
-and-io/ethernet-controller-e810/technical-library.html
Application Device Queues (ADQ)
-------------------------------
Application Device Queues (ADQ) allow you to dedicate one or more queues to a
specific application. This can reduce latency for the specified application,
and allow Tx traffic to be rate limited per application.
The ADQ information contained here is specific to the ice driver. For more
details, refer to the E810 ADQ Configuration Guide at:
https://cdrdv2.intel.com/v1/dl/getContent/609008
Requirements:
- Kernel version 4.19.58 or later
- Operating system: Red Hat* Enterprise Linux* 7.5+ or SUSE* Linux Enterprise
Server* 12+
- The sch_mqprio, act_mirred and cls_flower modules must be loaded. For example:
# modprobe sch_mqprio
# modprobe act_mirred
# modprove cls_flower
- The latest version of iproute2
# cd iproute2
# ./configure
# make DESTDIR=/opt/iproute2 install
- The latest ice driver and NVM image (Note: You must compile the ice driver
with the ADQ flag as shown in the "Building and Installation" section.)
When ADQ is enabled:
- You cannot change RSS parameters, the number of queues, or the MAC address in
the PF or VF. Delete the ADQ configuration before changing these settings.
- The driver supports subnet masks for IP addresses in the PF and VF. When you
add a subnet mask filter, the driver forwards packets to the ADQ VSI instead of
the main VSI.
- When the PF adds or deletes a port VLAN filter for the VF, it will extend to
all the VSIs within that VF.
Known issues:
- The latest RHEL and SLES distros have kernels with back-ported support for
ADQ. For all other Linux distributions, you must use the latest out-of-tree
driver to use ADQ.
- If the application stalls, the application-specific queues may stall for up
to two seconds. Configuring only one application per Traffic Class (TC) channel
may resolve the issue.
- DCB and ADQ cannot coexist. A switch with DCB enabled might remove the ADQ
configuration from the device. To resolve the issue, do not enable DCB on the
switch ports being used for ADQ. You must disable LLDP on the interface and
stop the firmware LLDP agent using the following command:
# ethtool --set-priv-flags <ethX> fw-lldp-agent off
- MACVLAN offloads and ADQ are mutually exclusive. System instability may occur
if you enable l2-fwd-offload and then set up ADQ, or if you set up ADQ and then
enable l2-fwd-offload.
- Commands such as 'tc qdisc add' and 'ethtool -L' will cause the driver to
close the associated RDMA interface and reopen it. This will disrupt RDMA
traffic for 3-5 seconds until the RDMA interface is available again for
traffic.
- Commands such as 'tc qdisc add' and 'ethtool -L' will clear other tuning
settings such as interrupt affinity. These tuning settings will need to be
reapplied. When the number of queues are increased using 'ethtool -L', the new
queues will have the same interrupt moderation settings as queue 0 (i.e., Tx
queue 0 for new Tx queues and Rx queue 0 for new Rx queues). You can change
this using the ethtool per-queue coalesce commands.
- TC filters may not get offloaded in hardware if you apply them immediately
after issuing the 'tc qdisc add' command. We recommend you wait 5 seconds after
issuing 'tc qdisc add' before adding TC filters. Dmesg will report the error if
TC filters fail to add properly.
To set up the adapter for ADQ, where <ethX> is the interface in use:
1. Reload the ice driver to remove any previous TC configuration:
# modprobe -r ice
# modprobe ice
2. Enable hardware TC offload on the interface:
# ethtool -K <ethX> hw-tc-offload on
3. Disable LLDP on the interface, if it isn't already:
# ethtool --set-priv-flags <ethX> fw-lldp-agent off
4. Verify settings:
# ethtool -k <ethX> | grep "hw-tc"
# ethtool --show-priv-flags <ethX>
Example output:
Private flags for p1p1:
link-down-on-close : off
fw-lldp-agent : off
channel-inline-flow-director : off
channel-pkt-inspect-optimize : on
To create traffic classes (TCs) on the interface:
NOTE: Run all TC commands from the ../iproute2/tc/ directory.
1. Use the tc command to create traffic classes. You can create a maximum of
16 TCs per interface.
# tc qdisc add dev <ethX> root mqprio num_tc <tcs> map <priorities>
queues <count1@offset1 ...> hw 1 mode channel shaper bw_rlimit
min_rate <min_rate1 ...> max_rate <max_rate1 ...>
Where:
num_tc <tcs>: The number of TCs to use.
map <priorities>: The map of priorities to TCs. You can map up to
16 priorities to TCs.
queues <count1@offset1 ...>: For each TC, <num queues>@<offset>. The max
total number of queues for all TCs is the number of cores.
hw 1 mode channel: 'channel' with 'hw' set to 1 is a new hardware offload
mode in mqprio that makes full use of the mqprio options, the TCs,
the queue configurations, and the QoS parameters.
shaper bw_rlimit: For each TC, sets the minimum and maximum bandwidth
rates. The totals must be equal to or less than the port speed. This
parameter is optional and is required only to set up the Tx rates.
min_rate <min_rate1>: Sets the minimum bandwidth rate limit for each TC.
max_rate <max_rate1 ...>: Sets the maximum bandwidth rate limit for each
TC. You can set a min and max rate together.
NOTE: See the mqprio man page and the examples below for more information.
2. Verify the bandwidth limit using network monitoring tools such as ifstat or
sar -n DEV [interval] [number of samples]
NOTE: Setting up channels via ethtool (ethtool -L) is not supported when the
TCs are configured using mqprio.
3. Enable hardware TC offload on the interface:
# ethtool -K <ethX> hw-tc-offload on
4. Apply TCs to ingress (Rx) flow of the interface:
# tc qdisc add dev <ethX> ingress
EXAMPLES:
See the tc and tc-flower man pages for more information on traffic control and
TC flower filters.
- To set up two TCs (tc0 and tc1), with 16 queues each, priorities 0-3 for
tc0 and 4-7 for tc1, and max Tx rate set to 1Gbit for tc0 and 3Gbit for tc1:
# tc qdisc add dev ens4f0 root mqprio num_tc 2 map 0 0 0 0 1 1 1 1 queues
16@0 16@16 hw 1 mode channel shaper bw_rlimit max_rate 1Gbit 3Gbit
Where:
map 0 0 0 0 1 1 1 1: Sets priorities 0-3 to use tc0 and 4-7 to use tc1
queues 16@0 16@16: Assigns 16 queues to tc0 at offset 0 and 16 queues
to tc1 at offset 16
- To set a minimum rate for a TC:
# tc qdisc add dev ens4f0 root mqprio num_tc 2 map 0 0 0 0 1 1 1 1 queues
4@0 8@4 hw 1 mode channel shaper bw_rlimit min_rate 25Gbit 50Gbit
- To set a maximum data rate for a TC:
# tc qdisc add dev ens4f0 root mqprio num_tc 2 map 0 0 0 0 1 1 1 1 queues
4@0 8@4 hw 1 mode channel shaper bw_rlimit max_rate 25Gbit 50Gbit
- To set both minimum and maximum data rates together:
# tc qdisc add dev ens4f0 root mqprio num_tc 2 map 0 0 0 0 1 1 1 1 queues
4@0 8@4 hw 1 mode channel shaper bw_rlimit min_rate 10Gbit 20Gbit
max_rate 25Gbit 50Gbit
- To configure TCP TC filters, where:
protocol: Encapsulation protocol (valid options are IP and 802.1Q).
prio: Priority.
flower: Flow-based traffic control filter.
dst_ip: IP address of the device.
ip_proto: IP protocol to use (TCP or UDP).
dst_port: Destination port.
src_port: Source port.
skip_sw: Flag to add the rule only in hardware.
hw_tc <tc>: Route incoming traffic flow to this hardware TC. The TC count
starts at 0. For example, hw_tc 1 indicates that the filter
is on the second TC.
- TCP: Destination IP + L4 Destination Port
To route TCP traffic with a matching destination IP address and destination
port to the given TC:
# tc filter add dev <ethX> protocol ip parent ffff: prio 1 flower dst_ip
<ip_address> ip_proto tcp dst_port <port_number> skip_sw hw_tc 1
- TCP: Destination IP + L4 Source Port
To route TCP traffic with a matching destination IP address and source port
to the given TC:
# tc filter add dev <ethX> protocol ip parent ffff: prio 1 flower dst_ip
<ip_address> ip_proto tcp src_port <port_number> skip_sw hw_tc 1
- To verify successful TC creation and traffic filtering, after filters are
created:
# tc qdisc show dev <ethX>
- To view all filters:
# tc filter show dev <ethX> parent ffff:
Multiple filters can be added to the device, using the same recipe (and
requires no additional recipe resources), either on the same interface or on
different interfaces. Each filter uses the same fields for matching, but can
have different match values.
# tc filter add dev <ethX> protocol ip ingress prio 1 flower ip_proto
tcp dst_port $app_port skip_sw hw_tc 1
For example:
# tc filter add dev <ethX> protocol ip ingress prio 1 flower ip_proto
tcp dst_port 5555 skip_sw hw_tc 1
Intel(R) Ethernet Flow Director
-------------------------------
The Intel Ethernet Flow Director performs the following tasks:
- Directs receive packets according to their flows to different queues
- Enables tight control on routing a flow in the platform
- Matches flows and CPU cores for flow affinity
NOTE: An included script (set_irq_affinity) automates setting the IRQ to CPU
affinity.
NOTE: This driver supports the following flow types:
- IPv4
- TCPv4
- UDPv4
- SCTPv4
- IPv6
- TCPv6
- UDPv6
- SCTPv6
Each flow type supports valid combinations of IP addresses (source or
destination) and UDP/TCP/SCTP ports (source and destination). You can supply
only a source IP address, a source IP address and a destination port, or any
combination of one or more of these four parameters.
NOTE: This driver allows you to filter traffic based on a user-defined flexible
two-byte pattern and offset by using the ethtool user-def and mask fields. Only
L3 and L4 flow types are supported for user-defined flexible filters. For a
given flow type, you must clear all Intel Ethernet Flow Director filters before
changing the input set (for that flow type).
NOTE: Flow Director filters impact only LAN traffic. RDMA filtering occurs
before Flow Director, so Flow Director filters will not impact RDMA.
The following table summarizes supported Intel Ethernet Flow Director features
across Intel(R) Ethernet controllers.
---------------------------------------------------------------------------
Feature 500 Series 700 Series 800 Series
===========================================================================
VF FLOW DIRECTOR Supported Routing to VF Not supported
not supported
---------------------------------------------------------------------------
IP ADDRESS RANGE Supported Not supported Field masking
FILTER
---------------------------------------------------------------------------
IPv6 SUPPORT Supported Supported Supported
---------------------------------------------------------------------------
CONFIGURABLE Configured Configured Configured
INPUT SET per port globally per port
---------------------------------------------------------------------------
ATR Supported Supported Not supported
---------------------------------------------------------------------------
FLEX BYTE FILTER Starts at Starts at Starts at
beginning beginning of beginning
of packet payload of packet
---------------------------------------------------------------------------
TUNNELED PACKETS Filter matches Filter matches Filter matches
outer header inner header inner header
---------------------------------------------------------------------------
Flow Director Filters
---------------------
Flow Director filters are used to direct traffic that matches specified
characteristics. They are enabled through ethtool's ntuple interface. To enable
or disable the Intel Ethernet Flow Director and these filters:
# ethtool -K <ethX> ntuple <off|on>
NOTE: When you disable ntuple filters, all the user programmed filters are
flushed from the driver cache and hardware. All needed filters must be re-added
when ntuple is re-enabled.
To display all of the active filters:
# ethtool -u <ethX>
To add a new filter:
# ethtool -U <ethX> flow-type <type> src-ip <ip> [m <ip_mask>] dst-ip <ip> [m
<ip_mask>] src-port <port> [m <port_mask>] dst-port <port> [m <port_mask>]
action <queue>
Where:
<ethX> - the Ethernet device to program
<type> - can be ip4, tcp4, udp4, sctp4, ip6, tcp6, udp6, sctp6
<ip> - the IP address to match on
<ip_mask> - the IPv4 address to mask on
NOTE: These filters use inverted masks.
<port> - the port number to match on
<port_mask> - the 16-bit integer for masking
NOTE: These filters use inverted masks.
<queue> - the queue to direct traffic toward (-1 discards the
matched traffic)
To delete a filter:
# ethtool -U <ethX> delete <N>
Where <N> is the filter ID displayed when printing all the active filters,
and may also have been specified using "loc <N>" when adding the filter.
EXAMPLES:
To add a filter that directs packet to queue 2:
# ethtool -U <ethX> flow-type tcp4 src-ip 192.168.10.1 dst-ip \
192.168.10.2 src-port 2000 dst-port 2001 action 2 [loc 1]
To set a filter using only the source and destination IP address:
# ethtool -U <ethX> flow-type tcp4 src-ip 192.168.10.1 dst-ip \
192.168.10.2 action 2 [loc 1]
To set a filter based on a user-defined pattern and offset:
# ethtool -U <ethX> flow-type tcp4 src-ip 192.168.10.1 dst-ip \
192.168.10.2 user-def 0x4FFFF action 2 [loc 1]
where the value of the user-def field contains the offset (4 bytes) and
the pattern (0xffff).
To match TCP traffic sent from 192.168.0.1, port 5300, directed to 192.168.0.5,
port 80, and then send it to queue 7:
# ethtool -U enp130s0 flow-type tcp4 src-ip 192.168.0.1 dst-ip 192.168.0.5
src-port 5300 dst-port 80 action 7
To add a TCPv4 filter with a partial mask for a source IP subnet:
# ethtool -U <ethX> flow-type tcp4 src-ip 192.168.0.0 m 0.255.255.255 dst-ip
192.168.5.12 src-port 12600 dst-port 31 action 12
NOTES:
For each flow-type, the programmed filters must all have the same matching
input set. For example, issuing the following two commands is acceptable:
# ethtool -U enp130s0 flow-type ip4 src-ip 192.168.0.1 src-port 5300 action 7
# ethtool -U enp130s0 flow-type ip4 src-ip 192.168.0.5 src-port 55 action 10
Issuing the next two commands, however, is not acceptable, since the first
specifies src-ip and the second specifies dst-ip:
# ethtool -U enp130s0 flow-type ip4 src-ip 192.168.0.1 src-port 5300 action 7
# ethtool -U enp130s0 flow-type ip4 dst-ip 192.168.0.5 src-port 55 action 10
The second command will fail with an error. You may program multiple filters
with the same fields, using different values, but, on one device, you may not
program two tcp4 filters with different matching fields.
The ice driver does not support matching on a subportion of a field, thus
partial mask fields are not supported.
Flex Byte Flow Director Filters
-------------------------------
The driver also supports matching user-defined data within the packet payload.
This flexible data is specified using the "user-def" field of the ethtool
command in the following way:
+----------------------------+--------------------------+
| 31 28 24 20 16 | 15 12 8 4 0 |
+----------------------------+--------------------------+
| offset into packet payload | 2 bytes of flexible data |
+----------------------------+--------------------------+
For example,
... user-def 0x4FFFF ...
tells the filter to look 4 bytes into the payload and match that value against
0xFFFF. The offset is based on the beginning of the payload, and not the
beginning of the packet. Thus
flow-type tcp4 ... user-def 0x8BEAF ...
would match TCP/IPv4 packets which have the value 0xBEAF 8 bytes into the
TCP/IPv4 payload.
Note that ICMP headers are parsed as 4 bytes of header and 4 bytes of payload.
Thus to match the first byte of the payload, you must actually add 4 bytes to
the offset. Also note that ip4 filters match both ICMP frames as well as raw
(unknown) ip4 frames, where the payload will be the L3 payload of the IP4 frame.
The maximum offset is 64. The hardware will only read up to 64 bytes of data
from the payload. The offset must be even because the flexible data is 2 bytes
long and must be aligned to byte 0 of the packet payload.
The user-defined flexible offset is also considered part of the input set and
cannot be programmed separately for multiple filters of the same type. However,
the flexible data is not part of the input set and multiple filters may use the
same offset but match against different data.
RSS Hash Flow
-------------
Allows you to set the hash bytes per flow type and any combination of one or
more options for Receive Side Scaling (RSS) hash byte configuration.
# ethtool -N <ethX> rx-flow-hash <type> <option>
Where <type> is:
tcp4 signifying TCP over IPv4
udp4 signifying UDP over IPv4
tcp6 signifying TCP over IPv6
udp6 signifying UDP over IPv6
And <option> is one or more of:
s Hash on the IP source address of the Rx packet.
d Hash on the IP destination address of the Rx packet.
f Hash on bytes 0 and 1 of the Layer 4 header of the Rx packet.
n Hash on bytes 2 and 3 of the Layer 4 header of the Rx packet.
Accelerated Receive Flow Steering (aRFS)
----------------------------------------
Devices based on the Intel(R) Ethernet 800 Series support Accelerated Receive
Flow Steering (aRFS) on the PF. aRFS is a load-balancing mechanism that allows
you to direct packets to the same CPU where an application is running or
consuming the packets in that flow.
NOTES:
- aRFS requires that ntuple filtering is enabled via ethtool.
- aRFS support is limited to the following packet types:
- TCP over IPv4 and IPv6
- UDP over IPv4 and IPv6
- Nonfragmented packets
- aRFS only supports Flow Director filters, which consist of the
source/destination IP addresses and source/destination ports.
- aRFS and ethtool's ntuple interface both use the device's Flow Director. aRFS
and ntuple features can coexist, but you may encounter unexpected results if
there's a conflict between aRFS and ntuple requests. See "Intel(R) Ethernet
Flow Director" for additional information.
To set up aRFS:
1. Enable the Intel Ethernet Flow Director and ntuple filters using ethtool.
# ethtool -K <ethX> ntuple on
2. Set up the number of entries in the global flow table. For example:
# NUM_RPS_ENTRIES=16384
# echo $NUM_RPS_ENTRIES > /proc/sys/net/core/rps_sock_flow_entries
3. Set up the number of entries in the per-queue flow table. For example:
# NUM_RX_QUEUES=64
# for file in /sys/class/net/$IFACE/queues/rx-*/rps_flow_cnt; do
# echo $(($NUM_RPS_ENTRIES/$NUM_RX_QUEUES)) > $file;
# done
4. Disable the IRQ balance daemon (this is only a temporary stop of the service
until the next reboot).
# systemctl stop irqbalance
5. Configure the interrupt affinity.
# set_irq_affinity <ethX>
To disable aRFS using ethtool:
# ethtool -K <ethX> ntuple off
NOTE: This command will disable ntuple filters and clear any aRFS filters in
software and hardware.
Example Use Case:
1. Set the server application on the desired CPU (e.g., CPU 4).
# taskset -c 4 netserver
2. Use netperf to route traffic from the client to CPU 4 on the server with
aRFS configured. This example uses TCP over IPv4.
# netperf -H <Host IPv4 Address> -t TCP_STREAM
Enabling Virtual Functions (VFs)
--------------------------------
Use sysfs to enable virtual functions (VF).
For example, you can create 4 VFs as follows:
# echo 4 > /sys/class/net/<ethX>/device/sriov_numvfs
To disable VFs, write 0 to the same file:
# echo 0 > /sys/class/net/<ethX>/device/sriov_numvfs
The maximum number of VFs for the ice driver is 256 total (all ports). To check
how many VFs each PF supports, use the following command:
# cat /sys/class/net/<ethX>/device/sriov_totalvfs
Note: You cannot use RDMA or SR-IOV when link aggregation (LAG)/bonding is
active, and vice versa. To enforce this, on kernels 4.5 and above, the driver
checks for this mutual exclusion. On kernels older than 4.5, the driver cannot
check for this exclusion and is unaware of bonding events.
Displaying VF Statistics on the PF
----------------------------------
Use the following command to display the statistics for the PF and its VFs:
# ip -s link show dev <ethX>
NOTE: The output of this command can be very large due to the maximum number of
possible VFs.
The PF driver will display a subset of the statistics for the PF and for all
VFs that are configured. The PF will always print a statistics block for each
of the possible VFs, and it will show zero for all unconfigured VFs.
Configuring VLAN Tagging on SR-IOV Enabled Adapter Ports
--------------------------------------------------------
To configure VLAN tagging for the ports on an SR-IOV enabled adapter, use the
following command. The VLAN configuration should be done before the VF driver
is loaded or the VM is booted. The VF is not aware of the VLAN tag being
inserted on transmit and removed on received frames (sometimes called "port
VLAN" mode).
# ip link set dev <ethX> vf <id> vlan <vlan id>
For example, the following will configure PF eth0 and the first VF on VLAN 10:
# ip link set dev eth0 vf 0 vlan 10
Enabling a VF link if the port is disconnected
----------------------------------------------
If the physical function (PF) link is down, you can force link up (from the
host PF) on any virtual functions (VF) bound to the PF. Note that this requires
kernel support (Red Hat kernel 3.10.0-327 or newer, upstream kernel 3.11.0 or
newer) and associated iproute2 user space support.
For example, to force link up on VF 0 bound to PF eth0:
# ip link set eth0 vf 0 state enable
Note: If the command does not work, it may not be supported by your system.
Setting the MAC Address for a VF
--------------------------------
To change the MAC address for the specified VF:
# ip link set <ethX> vf 0 mac <address>
For example:
# ip link set <ethX> vf 0 mac 00:01:02:03:04:05
This setting lasts until the PF is reloaded.
NOTE: For untrusted VFs, assigning a MAC address for a VF from the host will
disable any subsequent requests to change the MAC address from within the VM.
This is a security feature. The VM is not aware of this restriction, so if this
is attempted in the VM, it will trigger MDD events. Trusted VFs are allowed to
change the MAC address from within the VM.
Trusted VFs and VF Promiscuous Mode
-----------------------------------
This feature allows you to designate a particular VF as trusted and allows that
trusted VF to request selective promiscuous mode on the Physical Function (PF).
To set a VF as trusted or untrusted, enter the following command in the
Hypervisor:
# ip link set dev <ethX> vf 1 trust [on|off]
NOTE: It's important to set the VF to trusted before setting promiscuous mode.
If the VM is not trusted, the PF will ignore promiscuous mode requests from the
VF. If the VM becomes trusted after the VF driver is loaded, you must make a
new request to set the VF to promiscuous.
Once the VF is designated as trusted, use the following commands in the VM to
set the VF to promiscuous mode. For promiscuous all:
# ip link set <ethX> promisc on
Where <ethX> is a VF interface in the VM
For promiscuous Multicast:
# ip link set <ethX> allmulticast on
Where <ethX> is a VF interface in the VM
NOTE: By default, the ethtool private flag vf-true-promisc-support is set to
"off," meaning that promiscuous mode for the VF will be limited. To set the
promiscuous mode for the VF to true promiscuous and allow the VF to see all
ingress traffic, use the following command:
# ethtool --set-priv-flags <ethX> vf-true-promisc-support on
The vf-true-promisc-support private flag does not enable promiscuous mode;
rather, it designates which type of promiscuous mode (limited or true) you will
get when you enable promiscuous mode using the ip link commands above. Note
that this is a global setting that affects the entire device. However, the
vf-true-promisc-support private flag is only exposed to the first PF of the
device. The PF remains in limited promiscuous mode regardless of the
vf-true-promisc-support setting.
Next, add a VLAN interface on the VF interface. For example:
# ip link add link eth2 name eth2.100 type vlan id 100
Note that the order in which you set the VF to promiscuous mode and add the
VLAN interface does not matter (you can do either first). The result in this
example is that the VF will get all traffic that is tagged with VLAN 100.
Virtual Function (VF) Tx Rate Limit
-----------------------------------
Use the ip command to configure the maximum or minimum Tx rate limit for a VF
from the PF interface.
For example, to set a maximum Tx rate limit of 8000Mbps for VF 0:
# ip link set eth0 vf 0 max_tx_rate 8000
For example, to set a minimum Tx rate limit of 1000Mbps for VF 0:
# ip link set eth0 vf 0 min_tx_rate 1000
NOTE:
- If DCB or ADQ are enabled on a PF, you cannot set a minimum Tx rate on the
VFs associated with that PF.
- If both DCB and ADQ are disabled on a PF, then you can set a minimum Tx rate
on the VFs associated with that PF.
- If you set a minimum Tx rate limit on a PF for SR-IOV VFs and then apply a
DCB or ADQ configuration, the PF cannot guarantee the minimum Tx rate limits
for those VFs.
- If you set a minimum Tx rate on VFs across multiple ports that have an
aggregate bandwidth over 100Gbps, the PFs cannot guarantee the minimum Tx rate
set for the VFs.
Malicious Driver Detection (MDD) for VFs
----------------------------------------
Some Intel Ethernet devices use Malicious Driver Detection (MDD) to detect
malicious traffic from the VF and disable Tx/Rx queues or drop the offending
packet until a VF driver reset occurs. You can view MDD messages in the PF's
system log using the dmesg command.
- If the PF driver logs MDD events from the VF, confirm that the correct VF
driver is installed.
- To restore functionality, you can manually reload the VF or VM or enable
automatic VF resets.
- When automatic VF resets are enabled, the PF driver will immediately reset
the VF and reenable queues when it detects MDD events on the receive path.
- If automatic VF resets are disabled, the PF will not automatically reset the
VF when it detects MDD events.
To enable or disable automatic VF resets, use the following command:
# ethtool --set-priv-flags <ethX> mdd-auto-reset-vf on|off
MAC and VLAN Anti-Spoofing Feature for VFs
------------------------------------------
When a malicious driver on a Virtual Function (VF) interface attempts to send a
spoofed packet, it is dropped by the hardware and not transmitted.
NOTE: This feature can be disabled for a specific VF:
# ip link set <ethX> vf <vf id> spoofchk {off|on}
Switchdev mode
--------------
The PF driver supports legacy and switchdev eSwitch modes. Switchdev mode
allows the driver to create additional port representor netdevs that enable a
control plane running on the host to configure filters for the VFs and also
handle default/exception traffic from the uplink and the VFs.
The driver loads in legacy mode by default. You can configure eSwitch modes
independently per physical port using the devlink command. You can change
between eSwitch modes only if no VFs have been created. If SR-IOV is enabled
and VFs are bound to the PF, you must do the following before changing between
switchdev or legacy mode:
- Unload all VFs that were bound
- Set the number of VFs on the PF to zero
NOTE:
- ADQ, trusted VFs, and L2 forwarding are not supported in switchdev mode.
- Switchdev mode is not persistent across reboots or driver reloads.
To configure the device in switchdev mode, enter the following, where
<pci/0000:##:##.#> is the PCI address of the PF:
# devlink dev eswitch set <pci/0000:##:##.#> mode switchdev
For example:
# devlink dev eswitch set pci/0000:17:00.0 mode switchdev
To configure the device in legacy mode:
# devlink dev eswitch set <pci/0000:##:##.#> mode legacy
To check the current eSwitch mode:
# devlink dev eswitch show <pci/0000:##:##.#>
The following filters can be offloaded in switchdev mode:
- Supported filter conditions:
L2: EtherType, src/dst MAC addresses, VLAN ID, VLAN priority, VLAN TPID
L3: src/dst IP addresses (IPv4 and IPv6), IP protocol
L4: src/dst port
- Supported filter actions: redirect, drop
To offload TC filters to the hardware, you must enable hw-tc-offload on the VF
port representor (VF_PR). To enable hardware TC offload on the interface:
# ethtool -K <ethX> hw-tc-offload on
To verify settings:
# ethtool -k <ethX> | grep "hw-tc"
Switchdev mode supports the following ip link commands to configure the VF:
* mac - supported
* vlan, qos, proto - supported
* rate - supported but deprecated; use max_tx_rate instead
* max_tx_rate - supported
* min_tx_rate - supported
* spoofchk - supported
* query_rss - supported
* state - supported
* trust - not supported
* node_guid - supported
* port_guid - supported
Jumbo Frames
------------
Jumbo Frames support is enabled by changing the Maximum Transmission Unit (MTU)
to a value larger than the default value of 1500.
Use the ip command to increase the MTU size. For example, enter the following
where <ethX> is the interface number:
# ip link set mtu 9000 dev <ethX>
# ip link set up dev <ethX>
This setting is not saved across reboots.
Add 'MTU=9000' to the following file to make the setting change permanent:
/etc/sysconfig/network-scripts/ifcfg-<ethX> for RHEL
or
/etc/sysconfig/network/<config_file> for SLES
NOTE: The maximum MTU setting for jumbo frames is 9702. This corresponds to the
maximum jumbo frame size of 9728 bytes.
NOTE: This driver will attempt to use multiple page sized buffers to receive
each jumbo packet. This should help to avoid buffer starvation issues when
allocating receive packets.
NOTE: Packet loss may have a greater impact on throughput when you use jumbo
frames. If you observe a drop in performance after enabling jumbo frames,
enabling flow control may mitigate the issue.
Speed and Duplex Configuration
------------------------------
You cannot set speed, duplex, or autonegotiation settings using ethtool.
To see the speed configurations your device supports, run the following:
# ethtool <ethX>
To have your device advertise supported speeds, use the following:
# ethtool -s <ethX> advertise N
Where N is a bitmask of the desired speeds.
For example, to have your device advertise 10000baseSR Full, use:
# ethtool -s <ethX> advertise 0x80000000000
For more details, please refer to the ethtool man page.
Data Center Bridging (DCB)
--------------------------
NOTE:
The kernel assumes that TC0 is available, and will disable Priority Flow
Control (PFC) on the device if TC0 is not available. To fix this, ensure TC0 is
enabled when setting up DCB on your switch.
DCB is a configuration Quality of Service implementation in hardware. It uses
the VLAN priority tag (802.1p) to filter traffic. That means that there are 8
different priorities that traffic can be filtered into. It also enables
priority flow control (802.1Qbb) which can limit or eliminate the number of
dropped packets during network stress. Bandwidth can be allocated to each of
these priorities, which is enforced at the hardware level (802.1Qaz).
DCB is normally configured on the network using the DCBX protocol (802.1Qaz), a
specialization of LLDP (802.1AB). The ice driver supports the following
mutually exclusive variants of DCBX support:
1) Firmware-based LLDP Agent
2) Software-based LLDP Agent
In firmware-based mode, firmware intercepts all LLDP traffic and handles DCBX
negotiation transparently for the user. In this mode, the adapter operates in
"willing" DCBX mode, receiving DCB settings from the link partner (typically a
switch). The local user can only query the negotiated DCB configuration. For
information on configuring DCBX parameters on a switch, please consult the
switch manufacturer's documentation.
In software-based mode, LLDP traffic is forwarded to the network stack and user
space, where a software agent can handle it. In this mode, the adapter can
operate in either "willing" or "nonwilling" DCBX mode and DCB configuration can
be both queried and set locally. This mode requires the FW-based LLDP Agent to
be disabled.
NOTE:
- You can enable and disable the firmware-based LLDP Agent using an ethtool
private flag. Refer to the "FW-LLDP (Firmware Link Layer Discovery Protocol)"
section in this README for more information.
- In software-based DCBX mode, you can configure DCB parameters using software
LLDP/DCBX agents that interface with the Linux kernel's DCB Netlink API. We
recommend using OpenLLDP as the DCBX agent when running in software mode. For
more information, see the OpenLLDP man pages and
https://github.com/intel/openlldp.
- The driver implements the DCB netlink interface layer to allow the user space
to communicate with the driver and query DCB configuration for the port.
- iSCSI with DCB is not supported.
L3 QoS mode
-----------
The ice driver supports setting DSCP-based Layer 3 Quality of Service (L3 QoS)
in the PF driver. The driver initializes in L2 QoS mode. L3 QoS mode is:
- Automatically enabled when the first DSCP/ToS to TC mapping is defined
- Automatically disabled when the last DSCP/ToS to TC mapping is removed
The following is an example of how to map a DSCP/ToS to a TC:
# lldptool -T -i <ethX> -V APP app=<prio>,<sel>,<pid>
where:
<prio>: The TC assigned to the DSCP/ToS code point
<sel>: 5 for DSCP to TC mapping
<pid>: The DSCP/ToS code point
For example, to map packets containing DSCP value 63 to traffic class 0 on
interface eth0:
# lldptool -T -i eth0 -V APP app=63,5,0
To remove a mapping, use the following:
# lldptool -T -I <ethX> -V APP -d app=<prio>,<sel>,<pid>
To view the currently configured mappings, use the following:
# lldptool -t -i <ethX> -V APP -c
NOTE:
- You cannot enable L3 QoS mode if RDMA is active.
- L3 QoS mode is not available when FW-LLDP is enabled. You also cannot enable
FW-LLDP if L3 QoS mode is active. Disable FW-LLDP before switching to L3 QoS
mode. Refer to the "FW-LLDP (Firmware Link Layer Discovery Protocol)" section
in this README for more information on disabling FW-LLDP.
- Once a mapping has been submitted for a DSCP value, another mapping for that
value will not be accepted until the first one has been deleted.
FW-LLDP (Firmware Link Layer Discovery Protocol)
------------------------------------------------
Use ethtool to change FW-LLDP settings. The FW-LLDP setting is per port and
persists across boots.
To enable LLDP:
# ethtool --set-priv-flags <ethX> fw-lldp-agent on
To disable LLDP:
# ethtool --set-priv-flags <ethX> fw-lldp-agent off
To check the current LLDP setting:
# ethtool --show-priv-flags <ethX>
NOTE: You must enable the UEFI HII "LLDP Agent" attribute for this setting to
take effect. If "LLDP AGENT" is set to disabled, you cannot enable it from the
OS.
Flow Control
------------
Ethernet Flow Control (IEEE 802.3x) can be configured with ethtool to enable
receiving and transmitting pause frames for ice. When transmit is enabled,
pause frames are generated when the receive packet buffer crosses a predefined
threshold. When receive is enabled, the transmit unit will halt for the time
delay specified when a pause frame is received.
NOTE: You must have a flow control capable link partner.
Flow Control is disabled by default.
Use ethtool to change the flow control settings.
To enable or disable Rx or Tx Flow Control:
# ethtool -A <ethX> rx <on|off> tx <on|off>
Note: This command only enables or disables Flow Control if auto-negotiation is
disabled. If auto-negotiation is enabled, this command changes the parameters
used for auto-negotiation with the link partner.
Note: Flow Control auto-negotiation is part of link auto-negotiation. Depending
on your device, you may not be able to change the auto-negotiation setting.
NOTE:
- The ice driver requires flow control on both the port and link partner. If
flow control is disabled on one of the sides, the port may appear to hang on
heavy traffic.
- You may encounter issues with link-level flow control (LFC) after disabling
DCB. The LFC status may show as enabled but traffic is not paused. To resolve
this issue, disable and reenable LFC using ethtool:
# ethtool -A <ethX> rx off tx off
# ethtool -A <ethX> rx on tx on
NAPI
----
This driver supports NAPI (Rx polling mode).
For more information on NAPI, see
https://www.linuxfoundation.org/collaborate/workgroups/networking/napi
MACVLAN
-------
This driver supports MACVLAN. Kernel support for MACVLAN can be tested by
checking if the MACVLAN driver is loaded. You can run 'lsmod | grep macvlan' to
see if the MACVLAN driver is loaded or run 'modprobe macvlan' to try to load
the MACVLAN driver.
NOTE:
- In passthru mode, you can only set up one MACVLAN device. It will inherit the
MAC address of the underlying PF (Physical Function) device.
ice devices support L2 Forwarding Offload. This will offload the processing
required for L2 Forwarding from the system processors to the ice device.
Perform the following steps to enable L2 Forwarding Offload:
1. Enable L2 Forwarding offload:
# ethtool -K <ethX> l2-fwd-offload on
2. Create the MACVLAN netdevs and bind them to the PF.
3. Bring up/enable the MACVLAN netdevs.
NOTE: MACVLAN offloads and ADQ are mutually exclusive. System instability may
occur if you enable l2-fwd-offload and then set up ADQ, or if you set up ADQ
and then enable l2-fwd-offload.
IEEE 802.1ad (QinQ) Support
---------------------------
The IEEE 802.1ad standard, informally known as QinQ, allows for multiple VLAN
IDs within a single Ethernet frame. VLAN IDs are sometimes referred to as
"tags," and multiple VLAN IDs are thus referred to as a "tag stack." Tag stacks
allow L2 tunneling and the ability to separate traffic within a particular VLAN
ID, among other uses.
The following are examples of how to configure 802.1ad (QinQ):
# ip link add link eth0 eth0.24 type vlan proto 802.1ad id 24
# ip link add link eth0.24 eth0.24.371 type vlan proto 802.1Q id 371
Where "24" and "371" are example VLAN IDs.
NOTES:
- 802.1ad (QinQ) is supported in 3.19 and later kernels.
- VLAN protocols use the following EtherTypes:
802.1Q = EtherType 0x8100
802.1ad = EtherType 0x88A8
Double VLANs
------------
Devices based on the Intel(R) Ethernet 800 Series can process up to two VLANs
in a packet when all the following are installed:
- ice driver version 1.4.0 or later
- NVM version 2.4 or later
- ice DDP package version 1.3.21 or later
If you don't use the versions above, the only supported VLAN configuration is
single 802.1Q VLAN traffic.
When two VLAN tags are present in a packet, the outer VLAN tag can be either
802.1Q or 802.1ad. The inner VLAN tag must always be 802.1Q.
Note the following limitations:
- For each VF, the PF can only allow VLAN hardware offloads (insertion and
stripping) of one type, either 802.1Q or 802.1ad.
- You can't enable or disable outer or single 802.1Q or 802.1ad filtering
separately. They are either both on or both off.
- In SR-IOV mode, the VF may not receive all network traffic based on the inner
VLAN header when VF true promiscuous mode (vf-true-promisc-support) and double
VLANs are enabled.
To enable outer or single 802.1Q VLAN insertion and stripping and disable
802.1ad VLAN insertion and stripping:
# ethtool -K <ethX> rxvlan on txvlan on rx-vlan-stag-hw-parse off
tx-vlan-stag-hw-insert off
To enable outer or single 802.1ad VLAN insertion and stripping and disable
802.1Q VLAN insertion and stripping:
# ethtool -K <ethX> rxvlan off txvlan off rx-vlan-stag-hw-parse on
tx-vlan-stag-hw-insert on
To enable outer or single VLAN filtering:
# ethtool -K <ethX> rx-vlan-filter on rx-vlan-stag-filter on
To disable outer or single VLAN filtering:
# ethtool -K <ethX> rx-vlan-filter off rx-vlan-stag-filter off
Combining QinQ with SR-IOV VFs
------------------------------
We recommend you always configure a port VLAN for the VF from the PF. If a port
VLAN is not configured, the VF driver may only offload VLANs via software. The
PF allows all VLAN traffic to reach the VF, and the VF manages all VLAN traffic.
When the device is configured for double VLANs and the PF has configured a port
VLAN:
- The VF can only offload guest VLANs for 802.1Q traffic.
- The VF can only configure VLAN filtering rules for guest VLANs using 802.1Q
traffic.
However, when the device is configured for double VLANs and the PF has NOT
configured a port VLAN:
- You must use iavf driver version 4.1.0 or later to offload and filter VLANs.
- The PF turns on VLAN pruning and antispoof in the VF's VSI by default. The VF
will not transmit or receive any tagged traffic until the VF requests a VLAN
filter.
- The VF can offload (insert and strip) the outer VLAN tag of 802.1Q or 802.1ad
traffic.
- The VF can create filter rules for the outer VLAN tag of both 802.1Q and
802.1ad traffic.
If the PF does not support double VLANs, the VF can hardware offload single
802.1Q VLANs without a port VLAN.
When the PF is enabled for double VLANs, for iavf drivers before version 4.1.x:
- VLAN hardware offloads and filtering are supported only when the PF has
configured a port VLAN.
- VLAN filtering, insertion, and stripping will be software offloaded when no
port VLAN is configured.
To see VLAN filtering and offload capabilities, use the following command:
# ethtool -k <ethX> | grep vlan
IEEE 1588 Precision Time Protocol (PTP) Hardware Clock (PHC)
------------------------------------------------------------
Precision Time Protocol (PTP) is used to synchronize clocks in a computer
network. PTP support varies among Intel devices that support this driver. Use
'ethtool -T <ethX>' to get a definitive list of PTP capabilities supported by
the device.
Tunnel/Overlay Stateless Offloads
---------------------------------
Supported tunnels and overlays include VXLAN, GENEVE, and others depending on
hardware and software configuration. Stateless offloads are enabled by default.
To view the current state of all offloads:
# ethtool -k <ethX>
UDP Segmentation Offload
------------------------
Allows the adapter to offload transmit segmentation of UDP packets with
payloads up to 64K into valid Ethernet frames. Because the adapter hardware is
able to complete data segmentation much faster than operating system software,
this feature may improve transmission performance.
In addition, the adapter may use fewer CPU resources.
NOTE:
- UDP transmit segmentation offload requires Linux kernel 4.18 or later.
- The application sending UDP packets must support UDP segmentation offload.
To enable/disable UDP Segmentation Offload, issue the following command:
# ethtool -K <ethX> tx-udp-segmentation [off|on]
Using Devlink to update a device's NVM
--------------------------------------
When you update the NVM on some devices, the update may use the devlink
interface, rather than the ethtool interface. This will happen if the following
are true:
- You are updating an Intel Ethernet 800 Series device.
- Your system is running a distro that supports the "devlink dev flash" command.
- The firmware currently installed on the device supports it.
- The new NVM conforms to the correct PLDM format.
Most of the functionality and commands are the same with the following
exceptions:
- You cannot update a device in Recovery Mode. (To update a device in recovery
mode, you must download and install the Intel Ethernet driver set.)
- You cannot update the OROM or Netlist as a separate update, only as part of a
full NVM update.
- If you specified a preservation level of PRESERVE_ALL, the system will
immediately perform an EMPR reset after the NVM update.
On devices that support it, you can also use the devlink command line directly
to update the device NVM. However, we recommend that you use NVMUpdate.
# devlink dev flash pci/0000:3b:00.0 file filename.bin
Where :
- pci/0000:3b:00.0 – The device you wish to update. You can get a list of
devices with the "devlink dev info" command.
- filename.bin – The file that contains the new NVM image.
Firmware Logs
-------------
The ice driver allows you to generate firmware logs for supported categories of
events, to help debug issues with Intel Customer Support. Firmware logs are
enabled by default.
Firmware logs are printed to dmesg. The driver groups these events into
categories, called "modules." Supported modules include:
* 00000001 - General (Bit 0)
* 00000002 - Control (Bit 1)
* 00000004 - Link Management (Bit 2)
* 00000008 - Link Topology Detection (Bit 3)
* 00000010 - Link Control Technology (Bit 4)
* 00000020 - I2C (Bit 5)
* 00000040 - SDP (Bit 6)
* 00000080 - MDIO (Bit 7)
* 00000100 - Admin Queue (Bit 8)
* 00000200 - Host DMA (Bit 9)
* 00000400 - LLDP (Bit 10)
* 00000800 - DCBx (Bit 11)
* 00001000 - DCB (Bit 12)
* 00002000 - XLR (function-level resets; Bit 13)
* 00004000 - NVM (Bit 14)
* 00008000 - Authentication (Bit 15)
* 00010000 - VPD (Vital Product Data; Bit 16)
* 00020000 - IOSF (Intel On-Chip System Fabric, Bit 17)
* 00040000 - Parser (Bit 18)
* 00080000 - Switch (Bit 19)
* 00100000 - Scheduler (Bit 20)
* 00200000 - TX Queue Management (Bit 21)
* 00400000 - ACL (Access Control List; Bit 22)
* 00800000 - Post (Bit 23)
* 01000000 - Watchdog (Bit 24)
* 02000000 - Task Dispatcher (Bit 25)
* 04000000 - Manageability (Bit 26)
* 08000000 - SyncE (Bit 27)
* 10000000 - Health (Bit 28)
* 20000000 - Time Sync (Bit 29)
* 40000000 - PF Registration (Bit 30)
* 80000000 - Module Version (Bit 31)
You can change the verbosity level of the firmware logs. You can set only one
log level per module, and each level includes the verbosity levels lower than
it. For instance, setting the level to "normal" will also log warning and error
messages. Available verbosity levels are:
0 = none
1 = error
2 = warning
3 = normal
4 = verbose
Firmware logs can overrun the dmesg buffer. Before loading the driver, use the
following command to redirect dmesg to a file:
# dmesg -w > filename.log
Use a bitmap to set the desired verbosity level for the module(s). (NOTE: You
must have dynamic debug enabled in the kernel.) For example, to set all events
to log warning messages, use the following command:
# sudo insmod ice.ko dyndbg="+p" fwlog_events=0x0FFFFFFF fwlog_level=2
To log verbose, normal, warning, and error messages for the ACL (Bit 22),
Switch (Bit 19), and Parser (Bit 18) modules, for example, use the following:
# sudo insmod ice.ko dyndbg="+p" fwlog_events=0x4C0000 fwlog_level=4
To dump the firmware logging configuration to dmesg, use the following commands:
# echo dump fwlog > command
# dmesg
Performance Optimization
========================
Driver defaults are meant to fit a wide variety of workloads, but if further
optimization is required, we recommend experimenting with the following
settings.
IRQ to Adapter Queue Alignment
------------------------------
Pin the adapter's IRQs to specific cores by disabling the irqbalance service
and using the included set_irq_affinity script. Please see the script's help
text for further options.
- The following settings will distribute the IRQs across all the cores
evenly:
# scripts/set_irq_affinity -x all <interface1> , [ <interface2>, ... ]
- The following settings will distribute the IRQs across all the cores that
are local to the adapter (same NUMA node):
# scripts/set_irq_affinity -x local <interface1> ,[ <interface2>, ... ]
- For very CPU-intensive workloads, we recommend pinning the IRQs to all
cores.
Rx Descriptor Ring Size
-----------------------
To reduce the number of Rx packet discards, increase the number of Rx
descriptors for each Rx ring using ethtool.
- Check if the interface is dropping Rx packets due to buffers being full
(rx_dropped.nic can mean that there is no PCIe bandwidth):
# ethtool -S <ethX> | grep "rx_dropped"
- If the previous command shows drops on queues, it may help to increase
the number of descriptors using 'ethtool -G':
# ethtool -G <ethX> rx <N>
Where <N> is the desired number of ring entries/descriptors
This can provide temporary buffering for issues that create latency while
the CPUs process descriptors.
Interrupt Rate Limiting
-----------------------
This driver supports an adaptive interrupt throttle rate (ITR) mechanism that
is tuned for general workloads. The user can customize the interrupt rate
control for specific workloads, via ethtool, adjusting the number of
microseconds between interrupts.
To set the interrupt rate manually, you must disable adaptive mode:
# ethtool -C <ethX> adaptive-rx off adaptive-tx off
For lower CPU utilization:
- Disable adaptive ITR and lower Rx and Tx interrupts. The examples below
affect every queue of the specified interface.
- Setting rx-usecs and tx-usecs to 80 will limit interrupts to about
12,500 interrupts per second per queue:
# ethtool -C <ethX> adaptive-rx off adaptive-tx off rx-usecs 80
tx-usecs 80
For reduced latency:
- Disable adaptive ITR and ITR by setting rx-usecs and tx-usecs to 0
using ethtool:
# ethtool -C <ethX> adaptive-rx off adaptive-tx off rx-usecs 0
tx-usecs 0
Per-queue interrupt rate settings:
- The following examples are for queues 1 and 3, but you can adjust other
queues.
- To disable Rx adaptive ITR and set static Rx ITR to 10 microseconds or
about 100,000 interrupts/second, for queues 1 and 3:
# ethtool --per-queue <ethX> queue_mask 0xa --coalesce adaptive-rx off
rx-usecs 10
- To show the current coalesce settings for queues 1 and 3:
# ethtool --per-queue <ethX> queue_mask 0xa --show-coalesce
Bounding interrupt rates using rx-usecs-high:
- Valid Range: 0-236 (0=no limit)
The range of 0-236 microseconds provides an effective range of 4,237 to
250,000 interrupts per second. The value of rx-usecs-high can be set
independently of rx-usecs and tx-usecs in the same ethtool command, and is
also independent of the adaptive interrupt moderation algorithm. The
underlying hardware supports granularity in 4-microsecond intervals, so
adjacent values may result in the same interrupt rate.
- The following command would disable adaptive interrupt moderation, and allow
a maximum of 5 microseconds before indicating a receive or transmit was
complete. However, instead of resulting in as many as 200,000 interrupts per
second, it limits total interrupts per second to 50,000 via the rx-usecs-high
parameter.
# ethtool -C <ethX> adaptive-rx off adaptive-tx off rx-usecs-high 20
rx-usecs 5 tx-usecs 5
Virtualized Environments
------------------------
In addition to the other suggestions in this section, the following may be
helpful to optimize performance in VMs.
- Using the appropriate mechanism (vcpupin) in the VM, pin the CPUs to
individual LCPUs, making sure to use a set of CPUs included in the
device's local_cpulist: /sys/class/net/<ethX>/device/local_cpulist.
- Configure as many Rx/Tx queues in the VM as available. (See the iavf driver
documentation for the number of queues supported.) For example:
# ethtool -L <virt_interface> rx <max> tx <max>
Known Issues/Troubleshooting
============================
Receive Error counts may be higher than the actual packet error count
---------------------------------------------------------------------
When a packet is received with more than one error, two bad packets may be
reported. This affects all devices based on 10G, or faster, controllers.
Dynamic Debug
-------------
If you encounter unexpected issues during driver load, some of the most useful
information for developers to receive in a bug report can include driver
logging. This logging uses a kernel feature called Dynamic Debug, which is
generally enabled in most kernel configurations (CONFIG_DYNAMIC_DEBUG=y).
To load the driver with dynamic debug enabled, run modprobe with the dyndbg
parameter:
# modprobe ice dyndbg=+p
The driver will then load and print debugging information into the kernel log
(dmesg) and is usually logged into the system log viewable by journalctl or in
/var/log/messages. Saving this information to a file and attaching it to any
bug report can help shorten the reproduction and debugging time for a developer.
To enable dynamic debug during runtime operation of the driver, use this
command:
# echo "module ice +p" > /sys/kernel/debug/dynamic_debug/control
For more details, see the Dynamic Debug documentation included in the Linux
kernel instructions.
PF Message Queue Overflow
-------------------------
The device driver can detect some types of anomalous behavior. When it does, it
will log the VF MAC address and associated PF MAC address. Using this
information, you can check the virtual machine (VM) that is using the VF MAC
address to ensure that the VM is operating correctly.
'ethtool -S' does not display Tx/Rx packet statistics
-----------------------------------------------------
Issuing the command 'ethtool -S' does not display Tx/Rx packet statistics. This
is by convention. Use other tools (such as the 'ip' command) that display
standard netdev statistics such as Tx/Rx packet statistics.
Unexpected Issues when the device driver and DPDK share a device
----------------------------------------------------------------
Unexpected issues may result when an ice device is in multi driver mode and the
kernel driver and DPDK driver are sharing the device. This is because access to
the global NIC resources is not synchronized between multiple drivers. Any
change to the global NIC configuration (writing to a global register, setting
global configuration by AQ, or changing switch modes) will affect all ports and
drivers on the device. Loading DPDK with the "multi-driver" module parameter
may mitigate some of the issues.
Fiber optics and auto-negotiation
---------------------------------
Modules based on 100GBASE-SR4, active optical cable (AOC), and active copper
cable (ACC) do not support auto-negotiation per the IEEE specification. To
obtain link with these modules, you must turn off auto-negotiation on the link
partner's switch ports.
'ethtool -a' autonegotiate result may vary between drivers
----------------------------------------------------------
For kernel versions 4.6 or higher, 'ethtool -a' will show the advertised and
negotiated autoneg settings. For kernel versions below 4.6, ethtool will only
report the negotiated link status.
The issue is cosmetic and does not affect functionality. Installing the latest
ice driver and upgrading your kernel to version 4.6 or higher will resolve the
issue.
AF_XDP fails to allocate buffers
--------------------------------
On kernels older than 5.3, you may see an undesirable CPU load during packet
processing if you enable AF_XDP in native mode and the Rx ring size is larger
than the UMEM fill queue. This is due to a known issue in the kernel and was
fixed in 5.3. To address the issue, upgrade your kernel to 5.3 or newer.
SCTP checksum offloads aren't indicated on Geneve tunnel
--------------------------------------------------------
For SCTP traffic over a Geneve tunnel, the SCTP checksum isn't offloaded to the
device, even when tx-checksum-sctp is on. This is due to a limitation in the
Linux kernel. However, for Rx traffic, the SCTP checksum is verified if
rx-checksumming is on. For both Tx and Rx traffic, you can offload the outer
UDP checksum to the device.
CentOS 7.2 Issues
-----------------
The following issues are specific to CentOS 7.2. Upgrading to the latest
version of the operating system will resolve these issues.
- base-r-fec mode is supposed to be on by default. On CentOS 7.2, Ethtool
--show-priv-flags shows that it is off, instead of on.
- ethtool -m <ethX> does not display optical module information as expected.
- You cannot create an ipv6 Intel(R) Ethernet Flow Director rule. For example:
# ethtool -U p1p1 flow-type tcp6 src-ip 3001:1::2:1:1 dst-ip 3001:1::1:1:1
src-port 22 dst-port 23 action 10
Returns a bad syntax error.
Incorrect link speed reported on older VF drivers
-------------------------------------------------
Linux distributions with older iavf or i40evf drivers (including Red Hat
Enterprise Linux 8) may show an incorrect link speed on VF interfaces. This
issue is cosmetic and does not affect VF functionality. To resolve the issue,
download the latest iavf driver.
Older VF drivers on Intel Ethernet 800 Series adapters
------------------------------------------------------
Some Windows* VF drivers from Release 22.9 or older may encounter errors when
loaded on a PF based on the Intel Ethernet 800 Series on Linux KVM. You may see
errors and the VF may not load. This issue does not occur starting with the
following Windows VF drivers:
- v40e64, v40e65: Version 1.5.65.0 and newer
To resolve this issue, download and install the latest iavf driver.
'VF X failed opcode 24' error message in dmesg on host
------------------------------------------------------
With a Microsoft Windows Server 2019 guest machine running on a Linux host, you
may see 'VF <vf_number> failed opcode 24' error messages in dmesg on the host.
This error is benign and does not affect traffic. Installing the latest iavf
driver in the guest will resolve the issue.
Windows guest OSs on a Linux host may not pass traffic across VLANs
-------------------------------------------------------------------
The VF is not aware of the VLAN configuration if you use Load Balancing and
Failover (LBFO) to configure VLANs in a Windows guest. VLANs configured using
LBFO on a VF driver may result in failure to pass traffic.
SR-IOV virtual functions have identical MAC addresses
-----------------------------------------------------
When you create multiple SR-IOV virtual functions, the VFs may have identical
MAC addresses. Only one VF will pass traffic, and all traffic on other VFs with
identical MAC addresses will fail. This is related to the
"MACAddressPolicy=persistent" setting in
/usr/lib/systemd/network/99-default.link.
To resolve this issue, edit the /usr/lib/systemd/network/99-default.link file
and change the MACAddressPolicy line to "MACAddressPolicy=none". For more
information, see the systemd.link man page.
MDD events in dmesg when creating maximum number of VLANs on the VF
-------------------------------------------------------------------
When you create the maximum number of VLANs on the VF, you may see MDD events
in dmesg on the host. This is due to the asynchronous design of the iavf
driver. It always reports success to any VLAN requests, but the requests may
fail later. The guest OS could try to send traffic on a VLAN that is not
configured on the VF, which will cause a Malicious Driver Detection (MDD) event
in dmesg on the host.
This issue is cosmetic. You do not need to reload the PF driver.
'ip address' or 'ip link' command displays an error on a single-port NIC
with 245+ VFs
------------------------------------------------------------------------
When you use the 'ip address' or 'ip link' command on a Linux host configured
with 245 or more VFs on a single-port adapter, you may encounter a "Buffer too
small for object" error. This is due to a known issue in the iproute2 tools.
Please use ifconfig instead of iproute2. You can install ifconfig via the
net-tools-deprecated package.
Support
=======
For general information, go to the Intel support website at:
http://www.intel.com/support/
or the Intel Wired Networking project hosted by Sourceforge at:
http://sourceforge.net/projects/e1000
If an issue is identified with the released source code on a supported kernel
with a supported adapter, email the specific information related to the issue
to e1000-devel@lists.sf.net.
License
=======
This program is free software; you can redistribute it and/or modify it under
the terms and conditions of the GNU General Public License, version 2, as
published by the Free Software Foundation.
This program is distributed in the hope it will be useful, but WITHOUT ANY
WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
PARTICULAR PURPOSE. See the GNU General Public License for more details.
You should have received a copy of the GNU General Public License along with
this program; if not, write to the Free Software Foundation, Inc., 51 Franklin
St - Fifth Floor, Boston, MA 02110-1301 USA.
The full GNU General Public License is included in this distribution in the
file called "COPYING".
Copyright(c) 2017 - 2021 Intel Corporation.
Trademarks
==========
Intel is a trademark or registered trademark of Intel Corporation or its
subsidiaries in the United States and/or other countries.
* Other names and brands may be claimed as the property of others.
