[RFC,bpf-next,0/4] add partial rx hw csum offload support for XDP

Message

Lorenzo Bianconi May 28, 2021, 5:43 p.m. UTC

Enable xdp rx checksum offload support for CHECKSUM_UNNECESSARY use-case.
Introduce flag field in xdp_buff/xdp_frame in order to save the checksum
result from the NIC and have a fast access to it performing XDP_REDIRECT.
CHECKSUM_COMPLETE is not supported yet since it will require adding the
csum result to the xdp_metadata area.
Moreover flag field will be reused for xdp multi-buff support.
This series has been tested generating UDP traffic with pktgen and performing
a xdp_redirect from an ixgbe device to a remote CPUMAP entry. PPS results show
a negligible penalty respect to the baseline where the UDP checksum has been
disabled. More info about the test can be found here [0].

[0] https://github.com/xdp-project/xdp-project/blob/master/areas/core/xdp_frame01_checksum.org

Lorenzo Bianconi (4):
  net: xdp: introduce flags field in xdp_buff and xdp_frame
  mvneta: return csum computation result from mvneta_rx_csum
  net: mvneta: report csum result in xdp_buff
  net: xdp: update csum building the skb

 drivers/net/ethernet/marvell/mvneta.c | 27 ++++++++------------
 include/net/xdp.h                     | 36 +++++++++++++++++++++++++++
 net/core/xdp.c                        |  2 +-
 3 files changed, 48 insertions(+), 17 deletions(-)

Comments

David Ahern May 28, 2021, 6:01 p.m. UTC | #1

On 5/28/21 11:43 AM, Lorenzo Bianconi wrote:
> Enable xdp rx checksum offload support for CHECKSUM_UNNECESSARY use-case.
> Introduce flag field in xdp_buff/xdp_frame in order to save the checksum
> result from the NIC and have a fast access to it performing XDP_REDIRECT.
> CHECKSUM_COMPLETE is not supported yet since it will require adding the
> csum result to the xdp_metadata area.
> Moreover flag field will be reused for xdp multi-buff support.
> This series has been tested generating UDP traffic with pktgen and performing
> a xdp_redirect from an ixgbe device to a remote CPUMAP entry. PPS results show
> a negligible penalty respect to the baseline where the UDP checksum has been
> disabled. More info about the test can be found here [0].
> 
> [0] https://github.com/xdp-project/xdp-project/blob/master/areas/core/xdp_frame01_checksum.org
> 

For the wider audience, another example of the performance benefit of
this rather simple change is XDP_REDIRECT to VMs (and containers) for
TCP traffic which requires checksum.

The VM piece requires the change to the tun driver from my original
patch which passes the VIRTIO_NET_HDR_F_DATA_VALID flag for
CHECKSUM_UNNECESSARY:

https://github.com/dsahern/linux/commit/b6b4d4ef9562383d8b407a873d30082afdc1b89c

(I can send that followup after this set lands.)

Using ConnectX-5 and the mlx5 change in the above commit I was seeing at
least a 300k pps improvement depending on the test. e.g., from 2M pps to
2.3M for 1500 MTU packets or 1.3M to 1.6M for 3400 MTU.

Similar results for containers.

Right now the checksum validation done by H/W is lost on the redirect.
This change propagates the validation and avoids the csum_partial
overhead as the skb is processed.