Instead of a one-liner describing each vdev argument, add a description
and example for each. Move the information describing preferred busy
polling from the "Limitations" section to the "Options" section where it
is better placed. Also make general grammar improvements.
Signed-off-by: Ciara Loftus <ciara.loftus@intel.com>
Reviewed-by: Qi Zhang <qi.z.zhang@intel.com>
DPDK assumes that users only want AF_XDP socket (XSK) into zero copy
mode when the kernel supports it. However, sometimes kernel driver
doesn't support it well and copy mode is more stable and preferred.
This patch allows using devarg "-a xx:xx.x,force_copy=1" to force the
AF_XDP socket into copy mode.
Signed-off-by: Xiaoyun Li <xiaoyun.li@intel.com>
Reviewed-by: Andrew Rybchenko <andrew.rybchenko@oktetlabs.ru>
When an AF_XDP PMD is created without specifying the 'start_queue', the
default Rx queue associated with the socket will be Rx queue 0. A common
scenario encountered by users new to AF_XDP is that they create the
socket on queue 0 however their interface is configured with many more
queues. In this case, traffic might land on for example queue 18 which
means it will never reach the socket.
This commit updates the AF_XDP documentation with instructions on how to
configure the interface to ensure the traffic will land on queue 0 and
thus reach the socket successfully.
Signed-off-by: Ciara Loftus <ciara.loftus@intel.com>
Acked-by: Ferruh Yigit <ferruh.yigit@intel.com>
When libxdp is used, the LIBXDP_OBJECT_PATH environment variable must be
set to the location of where libxdp placed its bpf object files. This is
usually in /usr/local/lib/bpf or /usr/local/lib64/bpf. Failure to do so
will result in the PMD not initialising correctly as the bpf program is
not found. Document this requirement.
Also, mention that the following logs which are generated on application
launch can be ignored:
libbpf: elf: skipping unrecognized data section(7) .xdp_run_config
libbpf: elf: skipping unrecognized data section(8) xdp_metadata
Signed-off-by: Ciara Loftus <ciara.loftus@intel.com>
Acked-by: Ferruh Yigit <ferruh.yigit@intel.com>
Secondary process support had been disabled for the AF_XDP PMD because
there was no logic in place to share the AF_XDP socket file descriptors
between the processes. This commit introduces this logic using the IPC
APIs.
Rx and Tx are disabled in the secondary process due to memory mapping of
the AF_XDP rings being assigned by the kernel in the primary process only.
However other operations including retrieval of stats are permitted.
Signed-off-by: Ciara Loftus <ciara.loftus@intel.com>
Acked-by: Stephen Hemminger <stephen@networkplumber.org>
AF_XDP support is deprecated in libbpf since v0.7.0 [1]. The libxdp library
now provides the functionality which once was in libbpf and which the
AF_XDP PMD relies on. This commit updates the AF_XDP meson build to use the
libxdp library if a version >= v1.2.2 is available. If it is not available,
only versions of libbpf prior to v0.7.0 are allowed, as they still contain
the required AF_XDP functionality.
libbpf still remains a dependency even if libxdp is present, as we use
libbpf APIs for program loading.
The minimum required kernel version for libxdp for use with AF_XDP is v5.3.
For the library to be fully-featured, a kernel v5.10 or newer is
recommended. The full compatibility information can be found in the libxdp
README.
v1.2.2 of libxdp includes an important fix required for linking with DPDK
which is why this version or greater is required. Meson uses pkg-config to
verify the version of libxdp on the system, so it is necessary that the
library is discoverable using pkg-config in order for the PMD to use it. To
verify this, you can run: pkg-config --modversion libxdp
[1] https://github.com/libbpf/libbpf/commit/277846bc6c15
Signed-off-by: Ciara Loftus <ciara.loftus@intel.com>
Removing the use of driver following PMD as its unnecessary.
Cc: stable@dpdk.org
Signed-off-by: Sean Morrissey <sean.morrissey@intel.com>
Signed-off-by: Conor Fogarty <conor.fogarty@intel.com>
Acked-by: John McNamara <john.mcnamara@intel.com>
Reviewed-by: Conor Walsh <conor.walsh@intel.com>
Reviewed-by: Ferruh Yigit <ferruh.yigit@intel.com>
This commit introduces support for preferred busy polling
to the AF_XDP PMD. This feature aims to improve single-core
performance for AF_XDP sockets under heavy load.
A new vdev arg is introduced called 'busy_budget' whose default
value is 64. busy_budget is the value supplied to the kernel
with the SO_BUSY_POLL_BUDGET socket option and represents the
busy-polling NAPI budget. To set the budget to a different value
eg. 256:
--vdev=net_af_xdp0,iface=eth0,busy_budget=256
Preferred busy polling is enabled by default provided a kernel with
version >= v5.11 is in use. To disable it, set the budget to zero.
The following settings are also strongly recommended to be used in
conjunction with this feature:
echo 2 | sudo tee /sys/class/net/eth0/napi_defer_hard_irqs
echo 200000 | sudo tee /sys/class/net/eth0/gro_flush_timeout
.. where eth0 is the interface being used by the PMD.
Signed-off-by: Ciara Loftus <ciara.loftus@intel.com>
Reviewed-by: Ferruh Yigit <ferruh.yigit@intel.com>
AF_XDP will not work on 32-bit kernels before version 5.4.
Document this restriction in the driver guide.
Signed-off-by: Ciara Loftus <ciara.loftus@intel.com>
AF_XDP PMDs who wish to share a UMEM must have a unique context
(ctx) ie. netdev,qid tuple. For instance, the following will not
work since both PMDs' contexts are identical.
--vdev net_af_xdp0,iface=ens786f1,start_queue=0,shared_umem=1
--vdev net_af_xdp1,iface=ens786f1,start_queue=0,shared_umem=1
Supporting this scenario would require locks, which would impact
the performance of the more typical cases - xsks with different
netdev,qid tuples.
Fixes: 74b46340e2 ("net/af_xdp: support shared UMEM")
Signed-off-by: Ciara Loftus <ciara.loftus@intel.com>
The new 'xdp_prog=<string>' vdev arg allows the user to specify the path to
a custom XDP program to be set on the device, instead of the default libbpf
one. The program must have an XSK_MAP of name 'xsks_map' which will allow
for the redirection of some packets to userspace and thus the PMD, using
some criteria defined in the program. This can be useful for filtering
purposes, for example if we only want a subset of packets to reach
userspace or to drop or process a subset of packets in the kernel.
Note: a netdev may only load one program.
Signed-off-by: Ciara Loftus <ciara.loftus@intel.com>
Tested-by: Xuekun Hu <xuekun.hu@intel.com>
Kernel v5.10 will introduce the ability to efficiently share a UMEM
between AF_XDP sockets bound to different queue ids on the same or
different devices. This patch integrates that functionality into the AF_XDP
PMD.
A PMD will attempt to share a UMEM with others if the shared_umem=1 vdev
arg is set. UMEMs can only be shared across PMDs with the same mempool, up
to a limited number of PMDs goverened by the size of the given mempool.
Sharing UMEMs is not supported for non-zero-copy (aligned) mode.
The benefit of sharing UMEM across PMDs is a saving in memory due to not
having to register the UMEM multiple times. Throughput was measured to
remain within 2% of the default mode (not sharing UMEM).
A version of libbpf >= v0.2.0 is required and the appropriate pkg-config
file for libbpf must be installed such that meson can determine the
version.
Signed-off-by: Ciara Loftus <ciara.loftus@intel.com>
Explain how kernel driver RX buffer sizes affect the maximum
MTU size in practice.
Signed-off-by: Ciara Loftus <ciara.loftus@intel.com>
Reviewed-by: Xiaolong Ye <xiaolong.ye@intel.com>
This patch enables the unaligned chunks feature for AF_XDP which allows
chunks to be placed at arbitrary places in the umem, as opposed to them
being required to be aligned to 2k. This allows for DPDK application
mempools to be mapped directly into the umem and in turn enable zero copy
transfer between umem and the PMD.
This patch replaces the zero copy via external mbuf mechanism introduced
in commit e9ff8bb719 ("net/af_xdp: enable zero copy by external mbuf").
The pmd_zero copy vdev argument is also removed as now the PMD will
auto-detect presence of the unaligned chunks feature and enable it if so
and otherwise fall back to copy mode if not detected.
Signed-off-by: Ciara Loftus <ciara.loftus@intel.com>
Signed-off-by: Kevin Laatz <kevin.laatz@intel.com>
Reviewed-by: Xiaolong Ye <xiaolong.ye@intel.com>
queue= parameter does not exist.
It might have been the previous name of the queue_count parameter, but
anyway, the default value 1 for the number of queues works fine.
Fixes: f1debd77ef ("net/af_xdp: introduce AF_XDP PMD")
Cc: stable@dpdk.org
Signed-off-by: David Marchand <david.marchand@redhat.com>
Acked-by: Xiaolong Ye <xiaolong.ye@intel.com>
This patch enables need_wakeup flag for Tx and fill rings, when this
flag is set by the driver, it means that the userspace application has
to explicitly wake up the kernel Rx or kernel Tx processing by issuing
a syscall. Poll() can wake up both and sendto() or its alternatives
will wake up Tx processing only.
This feature is to provide efficient support for case that application
and driver executing on the same core.
Signed-off-by: Xiaolong Ye <xiaolong.ye@intel.com>
Reviewed-by: Ferruh Yigit <ferruh.yigit@intel.com>
This patch adds two parameters `start_queue` and `queue_count` to
specify the range of netdev queues used by AF_XDP pmd.
Signed-off-by: Xiaolong Ye <xiaolong.ye@intel.com>
Implement zero copy of af_xdp pmd through mbuf's external memory
mechanism to achieve high performance.
This patch also provides a new parameter "pmd_zero_copy" for user, so they
can choose to enable zero copy of af_xdp pmd or not.
To be clear, "zero copy" here is different from the "zero copy mode" of
AF_XDP, it is about zero copy between af_xdp umem and mbuf used in dpdk
application.
Suggested-by: Vipin Varghese <vipin.varghese@intel.com>
Suggested-by: Tummala Sivaprasad <sivaprasad.tummala@intel.com>
Suggested-by: Olivier Matz <olivier.matz@6wind.com>
Signed-off-by: Xiaolong Ye <xiaolong.ye@intel.com>
Add a new PMD driver for AF_XDP which is a proposed faster version of
AF_PACKET interface in Linux. More info about AF_XDP, please refer to [1]
[2].
This is the vanilla version PMD which just uses a raw buffer registered as
the umem.
[1] https://fosdem.org/2018/schedule/event/af_xdp/
[2] https://lwn.net/Articles/745934/
Signed-off-by: Xiaolong Ye <xiaolong.ye@intel.com>
Reviewed-by: Ferruh Yigit <ferruh.yigit@intel.com>
Acked-by: Luca Boccassi <bluca@debian.org>
Reviewed-by: Stephen Hemminger <stephen@networkplumber.org>
