Message-Id: <20180417200059.30154-1-willemdebruijn.kernel@gmail.com>
Date:   Tue, 17 Apr 2018 16:00:50 -0400
From:   Willem de Bruijn <willemdebruijn.kernel@...il.com>
To:     netdev@...r.kernel.org
Cc:     Willem de Bruijn <willemb@...gle.com>
Subject: [PATCH RFC net-next 00/11] udp gso 

From: Willem de Bruijn <willemb@...gle.com>

Segmentation offload reduces cycles/byte for large packets by
amortizing the cost of protocol stack traversal.

This patchset implements GSO for UDP. A process can concatenate and
submit multiple datagrams to the same destination in one send call
by setting socket option SOL_UDP/UDP_SEGMENT with the segment size,
or passing an analogous cmsg at send time.

The stack will send the entire large (up to network layer max size)
datagram through the protocol layer. At the GSO layer, it is broken
up in individual segments. All receive the same network layer header
and UDP src and dst port. All but the last segment have the same UDP
header, but the last may differ in length and checksum.

This initial patchset is RFC. A few open items

* MSG_MORE
  The feature requires UDP checksum offload, as without it the
  checksum + copy operation at send() time is likely cheaper than
  checksumming each segment in the GSO layer.

  UDP checksum offload is disabled with MSG_MORE. As a result, GSO
  only works in the lockless fast path.

  The patchset can be simplified if explicitly excluding MSG_MORE.
  For one, patch 1 can be dropped by passing ipcm to udp_send_skb
  instead of inet_cork.

* MSG_ZEROCOPY
  UDP zerocopy has been sent for review before. Completion
  notification cost exceeds the savings from copy avoidance for
  datagrams of regular MSS (< 1500B).

  UDP GSO enables building larger packets, at which point
  zerocopy becomes effective. Results with the current benchmark
  are not as great as from GSO itself, though that may say more
  about the benchmark. Either way, I do not intend to submit
  this separate feature as part of a final UDP GSO patchset.

* GSO_BY_FRAGS
  An alternative implementation that would allow non-uniform
  segment length is to use GSO_BY_FRAGS like SCTP. This would
  likely require MSG_MORE to build the list using multiple
  send calls (or one sendmmsg). The two approaches are not
  mutually-exclusive, so that could be a follow-up.

Initial results show a significant reduction in UDP cycles/byte.
See the main patch for more details and benchmark results.

        udp
          876 MB/s 14873 msg/s 624666 calls/s
            11,205,777,429      cycles
    
        udp gso
         2139 MB/s 36282 msg/s 36282 calls/s
            11,204,374,561      cycles


The patch set is broken down as follows:
- patch 1 is a prerequisite: code rearrangement, noop otherwise
- patch 2 is the core feature
- patch 3,4,6 are refinements
- patch 5 adds the cmsg interface
- patch 7 adds udp zerocopy
- patch 8..11 are tests

This idea was presented previously at netconf 2017-2
http://vger.kernel.org/netconf2017_files/rx_hardening_and_udp_gso.pdf

Known limitation:
  - The feature requires pacing and possibly a lower threshold on
    segment size to limit the number of segments that may be passed
    to the NIC at once.

  - Even when only accepting datagrams with CHECKSUM_PARTIAL, the
    segmentation layer must drop or fall back to software checksumming
    if the device cannot checksum the packet.
	    
    This can happen if a device advertises checksum offload in
    general, but removes it for this skb in .ndo_features_check.

Willem de Bruijn (11):
  udp: expose inet cork to udp
  udp: add gso
  udp: better wmem accounting on gso
  udp: paged allocation with gso
  udp: add gso segment cmsg
  udp: add gso support to virtual devices
  udp: zerocopy
  selftests: udp gso
  selftests: udp gso with connected sockets
  selftests: udp gso with corking
  selftests: udp gso benchmark

 include/linux/netdev_features.h               |   3 +
 include/linux/skbuff.h                        |  10 +
 include/linux/udp.h                           |   1 +
 include/net/inet_sock.h                       |   1 +
 include/net/ip.h                              |   3 +-
 include/net/ipv6.h                            |   2 +
 include/net/udp.h                             |   5 +
 include/uapi/linux/udp.h                      |   1 +
 net/core/skbuff.c                             |  14 +-
 net/core/sock.c                               |   5 +-
 net/ipv4/af_inet.c                            |   2 +-
 net/ipv4/ip_output.c                          |  63 +-
 net/ipv4/udp.c                                |  78 ++-
 net/ipv4/udp_offload.c                        |  63 ++
 net/ipv6/ip6_offload.c                        |   5 +-
 net/ipv6/ip6_output.c                         |  66 +-
 net/ipv6/udp.c                                |  29 +-
 net/ipv6/udp_offload.c                        |  14 +
 tools/testing/selftests/net/.gitignore        |   3 +
 tools/testing/selftests/net/Makefile          |   3 +-
 tools/testing/selftests/net/udpgso.c          | 621 ++++++++++++++++++
 tools/testing/selftests/net/udpgso.sh         |  31 +
 tools/testing/selftests/net/udpgso_bench.sh   |  74 +++
 tools/testing/selftests/net/udpgso_bench_rx.c | 265 ++++++++
 tools/testing/selftests/net/udpgso_bench_tx.c | 379 +++++++++++
 25 files changed, 1689 insertions(+), 52 deletions(-)
 create mode 100644 tools/testing/selftests/net/udpgso.c
 create mode 100755 tools/testing/selftests/net/udpgso.sh
 create mode 100755 tools/testing/selftests/net/udpgso_bench.sh
 create mode 100644 tools/testing/selftests/net/udpgso_bench_rx.c
 create mode 100644 tools/testing/selftests/net/udpgso_bench_tx.c

-- 
2.17.0.484.g0c8726318c-goog

Powered by blists - more mailing lists