[Intel-wired-lan] [next PATCH v3 00/15] TSO and checksum fixes for i40e

Alexander Duyck aduyck at mirantis.com
Mon Jan 25 05:16:05 UTC 2016


This patch set is meant to improve the performance and reliability of i40e
when it comes to performing TSO and Tx checksum offloads related to
tunnels.

I have tested it with a number of combinations of v4 over v6 and v6 over v4
for VXLANs.  With GRE I have only been able to test with v4 as the base as
it appears that offloads are currently not supported for GRE over IPv6.
With this patch set I resolved a number of issues and I am now able to
perform TSO for any of them as long as the outer UDP checksum is 0.  It
should also now be supported if the outer checksum is enabled in the case
of the XL722, though I cannot test it.

One additional item I have identified as an issue for the v2 patchset is
the fact that Rx checksums were not working for IPv6 based tunnels due to
the fact that the ports were not being registered.  From what I can tell
there was no reason for this so I have enabled IPv6 tunnels to register
their port numbers so that they can take full advantage of Rx offloads.

v2: Expanded coverage to include i40evf
    Enabled IPv6 tunnel support for Rx
    Moved XL722 support from generic checksum to TSO
v3: Fixed comments to match code in checksum path
    Dropped unused flag for checksum
    Dropped protocol parameter since it was unused in ATR function

Testing Hints:
I would recommend setting up a series of tunnels between two endpoints.
For VXLAN both v4 and v6 based tunnels can be used, with and without 
checksums.  For GRE a v4 tunnel can be used with and without checksum.

Significant performance improvements should be seen for IPv6 based tunnels,
or IPv4 tunnels passing IPv6 traffic with netperf.

The XL722 should no longer insert checksums for tunnels that do not have
udpcsum set.  This can be verified via tcpdump on the link partner.  XL722
should show significant performance improvements for transmit with tunnels
that have udpcsum with netperf TCP_STREAM testing.

The i40evf driver will show significant Tx performance improvements for
tunnels.  This can be verified with netperf.

---

Alexander Duyck (15):
      i40e/i40evf: Drop outer checksum offload that was not requested
      i40e/i40evf: Use u64 values instead of casting them in TSO function
      i40e/i40evf: Factor out L4 header and checksum from L3 bits in TSO path
      i40e/i40evf: Consolidate all header changes into TSO function
      i40e/i40evf: Replace header pointers with unions of pointers in Tx checksum path
      i40e/i40evf: Add support for IPv4 encapsulated in IPv6
      i40e/i40evf: Handle IPv6 extension headers in checksum offload
      i40e/i40evf: Do not write to descriptor unless we complete
      i40e/i40evf: Add exception handling for Tx checksum
      i40e/i40evf: Clean-up Rx packet checksum handling
      i40e/i40evf: Enable support for SKB_GSO_UDP_TUNNEL_CSUM
      i40e: Fix ATR in relation to tunnels
      i40e: Do not drop support for IPv6 VXLAN or GENEVE tunnels
      i40e: Update feature flags to reflect newly enabled features
      i40evf: Update feature flags to reflect newly enabled features


 drivers/net/ethernet/intel/i40e/i40e_main.c     |   30 +-
 drivers/net/ethernet/intel/i40e/i40e_txrx.c     |  385 ++++++++++++-----------
 drivers/net/ethernet/intel/i40e/i40e_txrx.h     |    2 
 drivers/net/ethernet/intel/i40evf/i40e_txrx.c   |  360 ++++++++++++----------
 drivers/net/ethernet/intel/i40evf/i40e_txrx.h   |    2 
 drivers/net/ethernet/intel/i40evf/i40evf_main.c |   23 +
 6 files changed, 424 insertions(+), 378 deletions(-)

--


More information about the Intel-wired-lan mailing list