XDP-bench is a benchmarking utility for exercising the different operation modes
of XDP. It is intended to be a simple program demonstrating the various
operating modes; these include dropping packets, hairpin forwarding (using the
XDP_TX return code), and redirection using the various in-kernel packet
redirection facilities.
The drop and TX modes support various options to control whether packet data is
touched (read or written) before being dropped or transmitted. The redirection
modes support using the simple ifindex-based bpf_redirect helper, the
bpf_redirect_map helper using a cpumap as its target, bpf_redirect_map using
a devmap as its target, and the devmap’s broadcast mode which allows redirecting
to multiple devices.
There is more information on the meaning of the output in both default (terse) and extended output mode, in the Output Format Description section below.
The syntax for running xdp-bench is:
Usage: xdp-bench COMMAND [options]
COMMAND can be one of:
drop - Drop all packets on an interface
pass - Pass all packets to the network stack
tx - Transmit packets back out on an interface (hairpin forwarding)
redirect - XDP redirect using the bpf_redirect() helper
redirect-cpu - XDP CPU redirect using BPF_MAP_TYPE_CPUMAP
redirect-map - XDP redirect using BPF_MAP_TYPE_DEVMAP
redirect-multi - XDP multi-redirect using BPF_MAP_TYPE_DEVMAP and the BPF_F_BROADCAST flag
xsk-drop - AF_XDP socket-based drop
xsk-tx - AF_XDP socket-based hairpin forwardingEach command, and its options are explained below. Or use xdp-bench COMMAND
--help to see the options for each command.
In this mode, xdp-bench installs an XDP program on an interface that simply
drops all packets. There are options to control what to do with the packet
before dropping it (touch the packet data or not), as well as which statistics
to gather. This is a basic benchmark for the baseline (best-case) performance of
XDP on an interface.
The syntax for the drop command is:
xdp-bench drop [options] <ifname>
Where <ifname> is the name of the interface the XDP program should be
installed on.
The supported options are:
Specify which operation should be taken on the packet before dropping it. The following actions are available:
no-touch - Drop the packet without touching the packet data
read-data - Read a field in the packet header before dropping
parse-ip - Parse the IP header field before dropping
swap-macs - Swap the source and destination MAC addresses before droppingWhether to touch the packet before dropping it can have a significant performance impact as this requires bringing packet data into the CPU cache (and flushing it back out if writing).
The default for this option is no-touch.
Specify which mechanism xdp-bench should use to load (and store) the packet data. The following modes are available:
dpa - Use traditional Direct Packet Access from the XDP program
load-bytes - Use the xdp_load_bytes() and xdp_store_bytes() helper functionsThis can be used to benchmark the various packet access modes supported by the kernel.
The default for this option is dpa.
If set, the XDP program will also gather statistics on which receive queue index each packet was received on. This is displayed in the extended output mode along with per-CPU data (which, depending on the hardware configuration may or may not be equivalent).
Set the polling interval for collecting all statistics and displaying them to the output. The unit of interval is in seconds.
Start xdp-bench in “extended” output mode. If not set, xdp-bench will start in “terse” mode. The output mode can be switched by hitting C-\ while the program is running. See also the Output Format Description section below.
Selects the XDP program mode (native or skb). Note that native XDP mode is the default, and loading the redirect program in skb manner is neither performant, nor recommended. However, this option is useful if the interface driver lacks native XDP support, or when simply testing the tool.
Enable verbose logging. Supply twice to enable verbose logging from the
underlying libxdp and libbpf libraries.
Show the application version and exit.
Display a summary of the available options
In this mode, xdp-bench installs an XDP program on an interface that passes
all packets to the network stack after processing them (returning XDP_PASS).
There are options to control what to do with the packet before passing it
(touch the packet data or not), as well as which statistics to gather. This is a
basic benchmark for the overhead of installing an XDP program on an interface
while still running the regular network stack.
The syntax for the pass command is:
xdp-bench pass [options] <ifname>
Where <ifname> is the name of the interface the XDP program should be
installed on.
The supported options are:
Specify which operation should be taken on the packet before passing it. The following actions are available:
no-touch - Pass the packet without touching the packet data
read-data - Read a field in the packet header before passing
parse-ip - Parse the IP header field before passing
swap-macs - Swap the source and destination MAC addresses before passingThe default for this option is no-touch.
Specify which mechanism xdp-bench should use to load (and store) the packet data. The following modes are available:
dpa - Use traditional Direct Packet Access from the XDP program
load-bytes - Use the xdp_load_bytes() and xdp_store_bytes() helper functionsThis can be used to benchmark the various packet access modes supported by the kernel.
The default for this option is dpa.
If set, the XDP program will also gather statistics on which receive queue index each packet was received on. This is displayed in the extended output mode along with per-CPU data (which, depending on the hardware configuration may or may not be equivalent).
Set the polling interval for collecting all statistics and displaying them to the output. The unit of interval is in seconds.
Start xdp-bench in “extended” output mode. If not set, xdp-bench will start in “terse” mode. The output mode can be switched by hitting C-\ while the program is running. See also the Output Format Description section below.
Selects the XDP program mode (native or skb). Note that native XDP mode is the default, and loading the redirect program in skb manner is neither performant, nor recommended. However, this option is useful if the interface driver lacks native XDP support, or when simply testing the tool.
Enable verbose logging. Supply twice to enable verbose logging from the
underlying libxdp and libbpf libraries.
Show the application version and exit.
Display a summary of the available options
In this mode, xdp-bench installs an XDP program on an interface that performs
so-called “hairpin forwarding”, which means each packet is transmitted back out
the same interface (using the XDP_TX return code).. There are options to
control what to do with the packet before transmitting it (touch the packet data
or not), as well as which statistics to gather.
The syntax for the tx command is:
xdp-bench tx [options] <ifname>
Where <ifname> is the name of the interface the XDP program should be
installed on.
The supported options are:
Specify which operation should be taken on the packet before transmitting it. The following actions are available:
no-touch - Transmit the packet without touching the packet data
read-data - Read a field in the packet header before transmitting
parse-ip - Parse the IP header field before transmitting
swap-macs - Swap the source and destination MAC addresses before transmittingTo allow the packet to be successfully transmitted back to the sender, the MAC addresses have to be swapped, so that the source MAC matches the network device. However, there is a performance overhead in doing swapping, so this option allows this function to be turned off.
The default for this option is swap-macs.
Specify which mechanism xdp-bench should use to load (and store) the packet data. The following modes are available:
dpa - Use traditional Direct Packet Access from the XDP program
load-bytes - Use the xdp_load_bytes() and xdp_store_bytes() helper functionsThis can be used to benchmark the various packet access modes supported by the kernel.
The default for this option is dpa.
If set, the XDP program will also gather statistics on which receive queue index each packet was received on. This is displayed in the extended output mode along with per-CPU data (which, depending on the hardware configuration may or may not be equivalent).
Set the polling interval for collecting all statistics and displaying them to the output. The unit of interval is in seconds.
Start xdp-bench in “extended” output mode. If not set, xdp-bench will start in “terse” mode. The output mode can be switched by hitting C-\ while the program is running. See also the Output Format Description section below.
Selects the XDP program mode (native or skb). Note that native XDP mode is the default, and loading the redirect program in skb manner is neither performant, nor recommended. However, this option is useful if the interface driver lacks native XDP support, or when simply testing the tool.
Enable verbose logging. Supply twice to enable verbose logging from the
underlying libxdp and libbpf libraries.
Show the application version and exit.
Display a summary of the available options
In this mode, xdp-bench sets up packet redirection between the two
interfaces supplied on the command line using the bpf_redirect BPF helper
triggered on packet reception on the ingress interface.
The syntax for the redirect command is:
xdp-bench redirect [options] <ifname_in> <ifname_out>
Where <ifname_in> is the name of the input interface from where packets will
be redirect to the output interface <ifname_out>.
The supported options are:
Specify which mechanism xdp-bench should use to load (and store) the packet data. The following modes are available:
dpa - Use traditional Direct Packet Access from the XDP program
load-bytes - Use the xdp_load_bytes() and xdp_store_bytes() helper functionsThis can be used to benchmark the various packet access modes supported by the kernel.
The default for this option is dpa.
Set the polling interval for collecting all statistics and displaying them to the output. The unit of interval is in seconds.
Enable statistics for successful redirection. This option comes with a per packet tracing overhead, for recording all successful redirections.
Start xdp-bench in “extended” output mode. If not set, xdp-bench will start in “terse” mode. The output mode can be switched by hitting C-\ while the program is running. See also the Output Format Description section below.
Selects the XDP program mode (native or skb). Note that native XDP mode is the default, and loading the redirect program in skb manner is neither performant, nor recommended. However, this option is useful if the interface driver lacks native XDP support, or when simply testing the tool.
Enable verbose logging. Supply twice to enable verbose logging from the
underlying libxdp and libbpf libraries.
Show the application version and exit.
Display a summary of the available options
In this mode, xdp-bench sets up packet redirection using the
bpf_redirect_map BPF helper triggered on packet reception on the ingress
interface, using a cpumap as its target. Hence, this tool can be used to
redirect packets on an interface from one CPU to another. In addition to this,
the tool then supports redirecting the packet to another output device when it
is processed on the target CPU.
The syntax for the redirect-cpu command is:
xdp-bench redirect-cpu [options] <ifname> -c 0 ... -c N
Where <ifname> is the name of the input interface from where packets will be
redirect to the target CPU list specified using -c.
The supported options are:
Specify a possible target CPU index. This option must be passed at least once,
and can be passed multiple times to specify a list of CPUs. Which CPU is chosen
for a given packet depends on the value of the --program-mode option,
described below.
Specify a program that embeds a predefined policy deciding how packets are redirected to different CPUs. The following options are available:
no-touch - Redirect without touching packet data
touch - Read packet data before redirecting
round-robin - Cycle between target CPUs in a round-robin fashion (for each packet)
l4-proto - Choose the target CPU based on the layer-4 protocol of packet
l4-filter - Like l4-proto, but drop UDP packets with destination port 9 (used by pktgen)
l4-hash - Use source and destination IP hashing to pick target CPU
l4-sport - Use modulo of source port to pick target CPU
l4-dport - Use modulo of destination port to pick target CPUThe no-touch and touch modes always redirect packets to the same CPU (the
first value supplied to --cpu). The round-robin and l4-hash modes
distribute packets between all the CPUs supplied as --cpu arguments, while
l4-proto and l4-filter send TCP and unrecognised packets to CPU index 0, UDP
packets to CPU index 1 and ICMP packets to CPU index 2 (where the index refers
to the order the actual CPUs are given on the command line).
The default for this option is l4-hash.
If this option is set, a separate program is installed into the cpumap, which
will be invoked on the remote CPU after the packet is processed there. The
action can be either drop or pass which will drop the packet or pass it to
the regular networking stack, respectively. Or it can be redirect, which will
cause the packet to be redirected to another interface and transmitted out that
interface on the remote CPU. If this option is set to redirect the target
device must be specified using --redirect-device.
The default for this option is disabled.
Specify the device to redirect the packet to when it is received on the target CPU.
Note that this option can only be specified with --remote-action redirect.
Set the queue size for the per-CPU cpumap ring buffer used for redirecting packets from multiple CPUs to one CPU. The default value is 2048 packets.
Stress the cpumap implementation by deallocating and reallocating the cpumap ring buffer on each polling interval.
Set the polling interval for collecting all statistics and displaying them to the output. The unit of interval is in seconds.
Enable statistics for successful redirection. This option comes with a per packet tracing overhead, for recording all successful redirections.
Start xdp-bench in “extended” output mode. If not set, xdp-bench will start in “terse” mode. The output mode can be switched by hitting C-\ while the program is running. See also the Output Format Description section below.
Selects the XDP program mode (native or skb). Note that native XDP mode is the default, and loading the redirect program in skb manner is neither performant, nor recommended. However, this option is useful if the interface driver lacks native XDP support, or when simply testing the tool.
Enable verbose logging. Supply twice to enable verbose logging from the
underlying libxdp and libbpf libraries.
Show the application version and exit.
Display a summary of the available options
In this mode, xdp-bench sets up packet redirection between two interfaces
supplied on the command line using the bpf_redirect_map() BPF helper triggered
on packet reception on the ingress interface, using a devmap as its target.
The syntax for the redirect-map command is:
xdp-bench redirect-map [options] <ifname_in> <ifname_out>
Where <ifname_in> is the name of the input interface from where packets will
be redirect to the output interface <ifname_out>.
The supported options are:
Load a program in the devmap entry used for redirection, so that it is invoked
after the packet is redirected to the target device, before it is transmitted
out of the output interface. The program can be selected via the egress-mode
option.
Set egress program to load:
forward - Update the packet data so its source MAC address matches the one of the destination interface.
drop - Drop packet.The default for this option is forward.
Set the polling interval for collecting all statistics and displaying them to the output. The unit of interval is in seconds.
Enable statistics for successful redirection. This option comes with a per packet tracing overhead, for recording all successful redirections.
Start xdp-bench in “extended” output mode. If not set, xdp-bench will start in “terse” mode. The output mode can be switched by hitting C-\ while the program is running. See also the Output Format Description section below.
Selects the XDP program mode (native or skb). Note that native XDP mode is the default, and loading the redirect program in skb manner is neither performant, nor recommended. However, this option is useful if the interface driver lacks native XDP support, or when simply testing the tool.
Enable verbose logging. Supply twice to enable verbose logging from the
underlying libxdp and libbpf libraries.
Show the application version and exit.
Display a summary of the available options
In this mode, xdp-bench sets up one-to-many packet redirection between
interfaces supplied on the command line, using the bpf_redirect_map BPF helper
triggered on packet reception on the ingress interface, using a devmap as its
target. The packet is broadcast to all output interfaces specified on the
command line, using devmap’s packet broadcast feature.
The syntax for the redirect-multi command is:
xdp-bench redirect-multi [options] <ifname_in> <ifname_out1> ... <ifname_outN>
Where <ifname_in> is the name of the input interface from where packets will
be redirect to one or many output interface(s).
The supported options are:
Load a program in the devmap entry used for redirection, so that it is invoked
after the packet is redirected to the target device, before it is transmitted
out of the output interface. The program can be selected via the egress-mode
option.
Set egress program to load:
forward - Update the packet data so its source MAC address matches the one of the destination interface.
drop - Drop packet.The default for this option is forward.
Set the polling interval for collecting all statistics and displaying them to the output. The unit of interval is in seconds.
Enable statistics for successful redirection. This option comes with a per packet tracing overhead, for recording all successful redirections.
Start xdp-bench in “extended” output mode. If not set, xdp-bench will start in “terse” mode. The output mode can be switched by hitting C-\ while the program is running. See also the Output Format Description section below.
Selects the XDP program mode (native or skb). Note that native XDP mode is the default, and loading the redirect program in skb manner is neither performant, nor recommended. However, this option is useful if the interface driver lacks native XDP support, or when simply testing the tool.
Enable verbose logging. Supply twice to enable verbose logging from the
underlying libxdp and libbpf libraries.
Show the application version and exit.
Display a summary of the available options
In these modes, xdp-bench drops or mirrors incoming packets (like when using the
drop or tx commands), but using an AF_XDP socket that passes the packets
directly from the XDP hook to userspace. These modes can be used to test and
benchmark AF_XDP sockets. The two commands differ only in the action taken -
the syntax and options are otherwise identical.
The code for these commands are derived from the xdpsock utility previously
part of the kernel sources, and more recently the bpf-examples repository. The
xsk-drop command corresponds to the --rxdrop option of xdpsock, and the
xsk-tx command corresponds to the --l2fwd command of xdpsock.
The syntax and options for the two commands are identical:
xdp-bench xsk-drop [options] <ifname>
or
xdp-bench xsk-tx [options] <ifname>
The supported options are:
Select the hardware receive queue to use for the test. The AF_XDP socket is
bound to this queue, which means that xdp-bench will process all packets
arriving on it. The default is to use queue 0.
Set the polling interval for collecting all statistics and displaying them to the output. The default interval is 2 seconds.
Set the size of the memory frames used to receive packets. Incoming packets will
always use a memory chunk of this size, regardless of the size of the packet on
the wire. The size must be a power of two, unless the --unaligned flag is
used. The default is 4096 bytes (which corresponds to the page size on most
systems).
Set the test duration, in seconds. The default is 0 (run forever).
Set the batch size for the receive loop. This is the maximum number of packets processed at once each time the loop runs. The default is 64 packets.
Enable printing of interrupt statistics for the interrupt number associated with
irq-string in /proc/interrupts. The string is used to find a matching IRQ
number by simple substring matching. No default (which means no IRQ statistics
are printed).
Use the poll() system call to get notifications for new packets, instead of
running in a loop constantly checking for new packets.
Don’t set the XDP_USE_NEED_WAKEUP flag when binding the AF_XDP socket. The
flag is set by default, and causes the kernel driver to occasionally go to sleep
and yield the CPU to the userspace application when receiving packets.
Support receiving unaligned packets, i.e., packets that are not aligned to the
beginning of a data frame. When set, xdp-bench will enable huge pages for the
userspace memory buffer used to receive packets. Setting this option enables
frame sizes that are not a power of two (see the --frame-size option).
Enable shared userspace memory across multiple AF_XDP sockets. When this
option is set, xdp-bench will create four sockets that share the same
userspace memory buffer, and packets will be sent to each socket in a
round-robin fashion.
Print extra packet statistics at each output interval.
Don’t print any statistics at each output interval.
Print additional application-level statistics at each output interval.
Set the SO_PREFER_BUSY_POLL socket option, which enables the application to
busy-poll without IRQs ever being enabled on the hardware queue being processed.
To make use of this option, the napi_defer_hard_irqs and gro_flush_timeout
sysctls should be set for the interface in question. For more information, see
https://docs.kernel.org/networking/napi.html#software-irq-coalescing.
Enable XDP fragments. This makes it possible to receive packets that are larger than the system page size, and requires kernel and driver support.
Select the copying mode for the packets. Set to one of auto, copy, or
zero-copy. The default is auto, which lets the kernel pick. Using zero-copy
requires kernel and driver support, and can only be used when attaching the XDP
program in native mode (see --attach-mode).
Choose the clock that will be used for all timer events in the application. The valid values are MONOTONIC, REALTIME, TAI and BOOTTIME, and the default is MONOTONIC.
Set the scheduler policy for the application before running. Valid values are
SCHED_OTHER and SCHED_FIFO, and the default is SCHED_OTHER.
Set the scheduler priority when using SCHED_FIFO. Defaults to 0.
Set the XDP program attach mode (native or skb). The default is native mode. Note that zero-copy can only be used in native mode.
Enable verbose logging. Supply twice to enable verbose logging from the
underlying libxdp and libbpf libraries.
Show the application version and exit.
Display a summary of the available options
By default, redirect success statistics are disabled, use --stats to enable.
The terse output mode is default, extended output mode can be activated using
the --extended command line option.
SIGQUIT (Ctrl + \) can be used to switch the mode dynamically at runtime.
Terse mode displays at most the following fields:
rx/s Number of packets received per second
redir/s Number of packets successfully redirected per second
err,drop/s Aggregated count of errors per second (including dropped packets when not using the drop command)
xmit/s Number of packets transmitted on the output device per secondExtended output mode displays at most the following fields:
FIELD DESCRIPTION
receive Displays the number of packets received and errors encountered
Whenever an error or packet drop occurs, details of per CPU error
and drop statistics will be expanded inline in terse mode.
pkt/s - Packets received per second
drop/s - Packets dropped per second
error/s - Errors encountered per second
redirect - Displays the number of packets successfully redirected
Errors encountered are expanded under redirect_err field
Note that passing -s to enable it has a per packet overhead
redir/s - Packets redirected successfully per second
redirect_err Displays the number of packets that failed redirection
The errno is expanded under this field with per CPU count
The recognized errors are:
EINVAL: Invalid redirection
ENETDOWN: Device being redirected to is down
EMSGSIZE: Packet length too large for device
EOPNOTSUPP: Operation not supported
ENOSPC: No space in ptr_ring of cpumap kthread
error/s - Packets that failed redirection per second
enqueue to cpu N Displays the number of packets enqueued to bulk queue of CPU N
Expands to cpu:FROM->N to display enqueue stats for each CPU enqueuing to CPU N
Received packets can be associated with the CPU redirect program is enqueuing
packets to.
pkt/s - Packets enqueued per second from other CPU to CPU N
drop/s - Packets dropped when trying to enqueue to CPU N
bulk-avg - Average number of packets processed for each event
kthread Displays the number of packets processed in CPUMAP kthread for each CPU
Packets consumed from ptr_ring in kthread, and its xdp_stats (after calling
CPUMAP bpf prog) are expanded below this. xdp_stats are expanded as a total and
then per-CPU to associate it to each CPU's pinned CPUMAP kthread.
pkt/s - Packets consumed per second from ptr_ring
drop/s - Packets dropped per second in kthread
sched - Number of times kthread called schedule()
xdp_stats (also expands to per-CPU counts)
pass/s - XDP_PASS count for CPUMAP program execution
drop/s - XDP_DROP count for CPUMAP program execution
redir/s - XDP_REDIRECT count for CPUMAP program execution
xdp_exception Displays xdp_exception tracepoint events
This can occur due to internal driver errors, unrecognized
XDP actions and due to explicit user trigger by use of XDP_ABORTED
Each action is expanded below this field with its count
hit/s - Number of times the tracepoint was hit per second
devmap_xmit Displays devmap_xmit tracepoint events
This tracepoint is invoked for successful transmissions on output
device but these statistics are not available for generic XDP mode,
hence they will be omitted from the output when using SKB mode
xmit/s - Number of packets that were transmitted per second
drop/s - Number of packets that failed transmissions per second
drv_err/s - Number of internal driver errors per second
bulk-avg - Average number of packets processed for each eventPlease report any bugs on Github: https://github.com/xdp-project/xdp-tools/issues
Earlier xdp-redirect tools were written by Jesper Dangaard Brouer and John Fastabend. They were then rewritten to support more features by Kumar Kartikeya Dwivedi, who also ported them to xdp-tools together with Toke Høiland-Jørgensen. This man page was written by Kumar Kartikeya Dwivedi and Toke Høiland-Jørgensen.