[v2,intel-next,0/4] XDP_TX improvements for ice

Message

Maciej Fijalkowski July 5, 2021, 4:43 p.m. UTC

Hi,

this is a second revision of a series around XDP_TX improvements for ice
driver. When compared to v1 (which can be found under [1]), two new
patches are introduced that are focused on improving the performance for
XDP_TX as Jussi reported that the numbers were pretty low on his side.
Furthermore the fallback path is now based on static branch, as
suggested by Toke on v1. This means that there's no further need for a
standalone net_device_ops that were serving the locked version of
ndo_xdp_xmit callback.

Idea from 2nd patch is borrowed from a joint work that was done against
OOT driver among with Sridhar Samudrala, Jesse Brandeburg and Piotr
Raczynski, where we working on fixing the scaling issues for Tx AF_XDP
ZC path.

Last but not least, with this series I observe the improvement of
performance by around 30%.

Thanks!
Maciej

[1] : https://lore.kernel.org/bpf/[e-mail]/

Maciej Fijalkowski (4):
  ice: unify xdp_rings accesses
  ice: optimize XDP_TX descriptor processing
  ice: do not create xdp_frame on XDP_TX
  ice: introduce XDP_TX fallback path

 drivers/net/ethernet/intel/ice/ice.h          |   3 +
 drivers/net/ethernet/intel/ice/ice_lib.c      |   4 +-
 drivers/net/ethernet/intel/ice/ice_main.c     |  45 ++++++-
 drivers/net/ethernet/intel/ice/ice_txrx.c     | 123 +++++++++++++++---
 drivers/net/ethernet/intel/ice/ice_txrx.h     |   6 +-
 drivers/net/ethernet/intel/ice/ice_txrx_lib.c |  29 +++--
 drivers/net/ethernet/intel/ice/ice_txrx_lib.h |  16 +++
 7 files changed, 190 insertions(+), 36 deletions(-)

Comments

Toke Høiland-Jørgensen July 5, 2021, 5:06 p.m. UTC | #1

Maciej Fijalkowski <[e-mail]> writes:

> Hi,
>
> this is a second revision of a series around XDP_TX improvements for ice
> driver. When compared to v1 (which can be found under [1]), two new
> patches are introduced that are focused on improving the performance for
> XDP_TX as Jussi reported that the numbers were pretty low on his side.
> Furthermore the fallback path is now based on static branch, as
> suggested by Toke on v1. This means that there's no further need for a
> standalone net_device_ops that were serving the locked version of
> ndo_xdp_xmit callback.
>
> Idea from 2nd patch is borrowed from a joint work that was done against
> OOT driver among with Sridhar Samudrala, Jesse Brandeburg and Piotr
> Raczynski, where we working on fixing the scaling issues for Tx AF_XDP
> ZC path.
>
> Last but not least, with this series I observe the improvement of
> performance by around 30%.

Wow, "but not least" indeed! :D
You can't just drop that at the end like that! I want details -
whaddyamean, "by around 30%"? In all cases? Only for TX? Gimme numbers! :)

-Toke