Current kernel allows offloading CRC32c computation when SCTP packets
are generated, setting skb->ip_summed to CHECKSUM_PARTIAL, if the
underlying device features have NETIF_F_SCTP_CRC set. However, after these
packets are forwarded, they may land on a device where CRC32c offloading is
not available: as a consequence, transmission is done with wrong CRC32c.
It's not possible to use sctp_compte_cksum() in the forwarding path
and in most drivers, because it needs symbols exported by libcrc32c module.
Patch 1 and 2 of this series try to solve this problem, introducing a new
helper function, namely skb_crc32c_csum_help(), that can be used to resolve
CHECKSUM_PARTIAL when crc32c is needed instead of Internet Checksum.
Currently, we need to parse the packet headers to understand what algorithm
is needed to resolve CHECKSUM_PARTIAL. We can speedup things by storing
this information in the skb metadata, and use it to call an appropriate
helper (skb_checksum_help or skb_crc32c_csum_help), or leave the packet
unmodified when the NIC is able to offload the checksum computation.
Patch 3 deprecates skb->csum_bad to free one bit in skb metadata; patch 4
introduces skb->csum_not_inet, providing skb with an indication on the
algorithm needed to resolve CHECKSUM_PARTIAL.
Patch 5 and 6 fix the kernel forwarding path and openvswitch datapath,
where skb_checksum_help was unconditionally called to resolve CHECKSUM_PARTIAL,
thus generating wrong CRC32c in forwarded SCTP packets.
Finally, patch 7 updates documentation to provide a better description of
possible values of skb->ip_summed.
Some further work is still possible:
* drivers that parse the packet header to correctly resolve CHECKSUM_PARTIAL
(e.g. ixgbe_tx_csum()) can benefit from testing skb->csum_not_inet to avoid
calling ip_hdr(skb)->protocol or ixgbe_ipv6_csum_is_sctp(skb).
* drivers that call skb_checksum_help() to resolve CHECKSUM_PARTIAL can
call skb_csum_hwoffload_help to avoid corrupting SCTP packets.
changes since RFCv4:
- patch 2/7: use WARN_ON_ONCE() instead of BUG_ON(), and avoid computing
CRC32c on the error path.
- patch 3/7: don't invert tests on the values of same_flow and
NAPI_GRO_CB(skb)->flush in dev_gro_receive(), it's useless and it breaks
GRO functionality as reported by kernel test robot.
Davide Caratti (7):
skbuff: add stub to help computing crc32c on SCTP packets
net: introduce skb_crc32c_csum_help
sk_buff: remove support for csum_bad in sk_buff
net: use skb->csum_not_inet to identify packets needing crc32c
net: more accurate checksumming in validate_xmit_skb()
openvswitch: more accurate checksumming in queue_userspace_packet()
sk_buff.h: improve description of CHECKSUM_{COMPLETE,UNNECESSARY}
Documentation/networking/checksum-offloads.txt | 11 +++--
drivers/net/ethernet/aquantia/atlantic/aq_ring.c | 2 +-
include/linux/netdevice.h | 8 ++--
include/linux/skbuff.h | 58 +++++++++--------------
net/bridge/netfilter/nft_reject_bridge.c | 5 +-
net/core/dev.c | 59 ++++++++++++++++++++++--
net/core/skbuff.c | 24 ++++++++++
net/ipv4/netfilter/nf_reject_ipv4.c | 2 +-
net/ipv6/netfilter/nf_reject_ipv6.c | 3 --
net/openvswitch/datapath.c | 2 +-
net/sched/act_csum.c | 1 +
net/sctp/offload.c | 8 ++++
net/sctp/output.c | 1 +
13 files changed, 126 insertions(+), 58 deletions(-)
--
2.7.4
--
To unsubscribe from this list: send the line "unsubscribe linux-sctp" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
