[PATCH 00/14] udp optimisation

From: Pavel Begunkov
Date: Mon Jan 10 2022 - 20:24:48 EST


A mainly UDP/IPv6 optimisation patch set. Zerocopy io_uring benchmark over
dummy netdev (CPU bound) gives 2068992 -> 2166481 tx/s, which is ~4.7% or
over 5% of net layer overhead. Should give similar results for small
packet non-zerocopy.

- 1/14 and 9/14 remove a get/put dst pair each, so saving 4 atomics per
corkless UDP send.
- Patches 3-8 optimise iflow handling, in particular removes one 88B
memset and one 88B copy.
- 10-14 are random improvements, which are not UDP-specific but also
beneficial to TCP and others.

Pavel Begunkov (14):
ipv6: optimise dst referencing
ipv6: shuffle up->pending AF_INET bits
ipv6: remove daddr temp buffer in __ip6_make_skb
ipv6: clean up cork setup/release
ipv6: don't zero cork's flowi after use
ipv6: pass full cork into __ip6_append_data()
ipv6: pass flow in ip6_make_skb together with cork
ipv6/udp: don't make extra copies of iflow
ipv6: hand dst refs to cork setup
skbuff: drop zero check from skb_zcopy_set
skbuff: drop null check from skb_zcopy
skbuff: optimise alloc_skb_with_frags()
net: inline part of skb_csum_hwoffload_help
net: inline sock_alloc_send_skb

include/linux/netdevice.h | 16 +++++-
include/linux/skbuff.h | 45 +++++++++++++---
include/net/ipv6.h | 2 +-
include/net/sock.h | 10 +++-
net/core/dev.c | 15 ++----
net/core/skbuff.c | 34 +++++-------
net/core/sock.c | 7 ---
net/ipv4/ip_output.c | 10 ++--
net/ipv4/tcp.c | 5 +-
net/ipv6/ip6_output.c | 105 +++++++++++++++++++++-----------------
net/ipv6/udp.c | 103 ++++++++++++++++++-------------------
11 files changed, 197 insertions(+), 155 deletions(-)

--
2.34.1