Skip to content

net: mitigate retpoline overhead

Paolo Abeni lkml netdev 3 versions – In review

Get series with b4
b4 am [email protected]

Versions newest first · linked by author and title

  1. v3 Latest4 patches 4 repliesIn review
  2. v24 patches 10 repliesSuperseded
  3. v14 patches 7 repliesSuperseded

Changes in v3 from the cover letter

- fix build error with CONFIG_IPV6=m

Patches in v3 No applies recorded yet

#Subject and filesSizeRepliesApplied
1/4indirect call wrappers: helpers to speed-up indirect calls of builtininclude/linux/indirect_call_wrapper.h+51 -00 replies
2/4net: use indirect call wrappers at GRO network layerinclude/net/inet_common.h net/core/dev.c net/ipv6/ip6_offload.c+18 -50 replies
3/4net: use indirect call wrappers at GRO transport layerinclude/net/inet_common.h net/ipv4/af_inet.c net/ipv4/tcp_offload.c +4 more+61 -150 replies
4/4udp: use indirect call wrappers for GRO socket lookupnet/ipv4/udp_offload.c+6 -20 replies

Reviews trailers per patch

No Reviewed-by, Acked-by or Tested-by trailers for v3 yet.

Read the v3 thread

Cover letter [PATCH net-next v3 0/4] net: mitigate retpoline overhead

The spectre v2 counter-measures, aka retpolines, are a source of measurable overhead[1]. We can partially address that when the function pointer refers to a builtin symbol resorting to a list of tests vs well-known builtin function and direct calls.

Experimental results show that replacing a single indirect call via retpoline with several branches and a direct call gives performance gains even when multiple branches are added - 5 or more, as reported in [2].

This may lead to some uglification around the indirect calls. In netconf 2018 Eric Dumazet described a technique to hide the most relevant part of the needed boilerplate with some macro help.

This series is a [re-]implementation of such idea, exposing the introduced helpers in a new header file. They are later leveraged to avoid the indirect call overhead in the GRO path, when possible.

Overall this gives > 10% performance improvement for UDP GRO benchmark and smaller but measurable for TCP syn flood.

The added infra can be used in follow-up patches to cope with retpoline overhead in other points of the networking stack (e.g. at the qdisc layer) and possibly even in other subsystems.

v2 -> v3:
- fix build error with CONFIG_IPV6=m

v1 -> v2:
- list explicitly the builtin function names in INDIRECT_CALL_*(), as suggested by Ed Cree
- expand the recipients list

rfc -> v1:
- use branch prediction hints, as suggested by Eric

[1] http://vger.kernel.org/netconf2018_files/PaoloAbeni_netconf2018.pdf
[2] https://linuxplumbersconf.org/event/2/contributions/99/attachments/98/117/lpc18_paper_af_xdp_perf-v2.pdf

Paolo Abeni (4):
indirect call wrappers: helpers to speed-up indirect calls of builtin
net: use indirect call wrappers at GRO network layer
net: use indirect call wrappers at GRO transport layer
udp: use indirect call wrappers for GRO socket lookup

 include/linux/indirect_call_wrapper.h | 51 +++++++++++++++++++++++++++
 include/net/inet_common.h             |  9 +++++
 net/core/dev.c                        | 15 ++++++--
 net/ipv4/af_inet.c                    | 13 +++++--
 net/ipv4/tcp_offload.c                |  6 ++--
 net/ipv4/udp_offload.c                | 15 +++++---
 net/ipv6/ip6_offload.c                | 35 +++++++++++++++---
 net/ipv6/tcpv6_offload.c              |  7 ++--
 net/ipv6/udp_offload.c                |  7 ++--
 9 files changed, 136 insertions(+), 22 deletions(-)
 create mode 100644 include/linux/indirect_call_wrapper.h
Signature · 1 line
2.19.2