[dpdk-dev] [PATCH v3 0/3] AF_XDP Preferred Busy Polling

[Ciara Loftus <ciara.loftus@intel.com>]

Single-core performance of AF_XDP at high loads can be poor because
a heavily loaded NAPI context will never enter or allow for busy-polling.

1C testpmd rxonly (both IRQs and PMD on core 0):
./dpdk-testpmd -l 0-1 --vdev=net_af_xdp0,iface=eth0 --main-lcore=1 -- \
--forward-mode=rxonly
0.088Mpps

In order to achieve decent performance at high loads, it is currently
recommended ensure the IRQs for the netdev queue and the core running
the PMD are different.

2C testpmd rxonly (IRQs on core 0, PMD on core 1):
./dpdk-testpmd -l 0-1 --vdev=net_af_xdp0,iface=eth0 --main-lcore=0 -- \
--forward-mode=rxonly
19.26Mpps

However using an extra core is of course not ideal. The SO_PREFER_BUSY_POLL
socket option was introduced in kernel v5.11 to help improve 1C performance.
See [1].

This series sets this socket option on xsks created with DPDK (ie. instances of
the AF_XDP PMD) unless explicitly disabled or not supported by the kernel. It
was chosen to be enabled by default in order to bring the AF_XDP PMD in line
with most other PMDs which execute on a single core.

The following system and netdev settings are recommended in conjunction with
busy polling:
echo 2 | sudo tee /sys/class/net/eth0/napi_defer_hard_irqs
echo 200000 | sudo tee /sys/class/net/eth0/gro_flush_timeout

Re-running the 1C test with busy polling support and the above settings:
./dpdk-testpmd -l 0-1 --vdev=net_af_xdp0,iface=eth0 --main-lcore=1 -- \
--forward-mode=rxonly
10.45Mpps

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 ie. the number of
packets the kernel will attempt to process in the netdev's NAPI context.

To set the busy budget to 256:
./dpdk-testpmd --vdev=net_af_xdp0,iface=eth0,busy_budget=256
14.06Mpps

If you still wish to run using 2 cores (one for PMD once for IRQs) it is
recommended to disable busy polling to achieve optimal 2C performance:
./dpdk-testpmd --vdev=net_af_xdp0,iface=eth0,busy_budget=0
19.09Mpps

v2->v3:
* Moved release notes update to correct location
* Changed busy_budget from uint32_t to int since this is the type expected
by setsockopt
* Validate busy_budget arg is <= UINT16_MAX during parse

v1->v2:
* Set batch size to default size of ring (2048)
* Split batches > 2048 into multiples of 2048 or less and process all
packets in the same manner that is done for other drivers eg. ixgbe:
http://code.dpdk.org/dpdk/v21.02/source/drivers/net/ixgbe/ixgbe_rxtx.c#L318
* Update commit log with reasoning behing batching changes
* Update release notes with note on busy polling support
* Fix return type for sycall_needed function when the wakeup flag is not
present
* Apprpriate log leveling
* Set default_*xportconf burst sizes to the default busy budget size (64)
* Detect support for busy polling via setsockopt instead of using the presence
of the flag

RFC->v1:
* Fixed behaviour of busy_budget=0
* Ensure we bail out any of the new setsockopts fail

[1] https://lwn.net/Articles/837010/


Ciara Loftus (3):
  net/af_xdp: allow bigger batch sizes
  net/af_xdp: Use recvfrom() instead of poll()
  net/af_xdp: preferred busy polling

 doc/guides/nics/af_xdp.rst             |  38 ++++-
 doc/guides/rel_notes/release_21_05.rst |   4 +
 drivers/net/af_xdp/compat.h            |  14 ++
 drivers/net/af_xdp/rte_eth_af_xdp.c    | 208 ++++++++++++++++++++++---
 4 files changed, 240 insertions(+), 24 deletions(-)

-- 
2.17.1