[Intel-wired-lan] [PATCH v5 0/2] i40e: support for XDP

[Alexander Duyck]

On Fri, May 19, 2017 at 12:08 AM, Björn Töpel <bjorn.topel at gmail.com> wrote:
> From: Björn Töpel <bjorn.topel at intel.com>
>
> This series adds XDP support for i40e-based NICs.
>
> The first patch wires up ndo_xdp and implements XDP_DROP semantics for
> all actions. The second patch adds egress support via the XDP_TX
> action.
>
> Performance numbers (40GbE port, 64B packets) for xdp1 and xdp2
> programs, from samples/bpf/:
>
>  IOMMU                      | xdp1      | xdp2
>  ---------------------------+-----------+-----------
>  iommu=off                  | 29.7 Mpps | 17.1 Mpps
>  iommu=pt intel_iommu=on    | 29.7 Mpps | 11.6 Mpps
>  iommu=on intel_iommu=on    | 21.8 Mpps |  3.7 Mpps

These numbers look pretty good. I wouldn't expect us to have much in
the way of performance with iommu enabled, and the iommu=off numbers
are about 20Gb/s for xdp1, and better than 10Gb/s for xdp2 so this is
a good starting point. I'm assuming this is a single queue throughput
test?

> Future improvements, not covered by the patches:
>   * Egress: Create the iova mappings upfront
>     (DMA_BIDIRECTIONAL/dma_sync_*), instead of creating a new iova
>     mapping in the transmit fast-path. This will improve performance
>     for the IOMMU-enabled case.

The problem with using DMA_BIDIRECTIONAL is that there are scenarios
where it makes DMA more expensive in general as we have to then push
the data every time we do a sync for CPU. If you take a look at the
swiotlb code it will give you an idea of what I am talking about.

Also when we start supporting redirection the DMA_BIDIRECTIONAL won't
be useful unless you want to map a buffer for multiple devices
simultaneously.

>   * Proper debugfs support.
>   * i40evf support.
>
> Thanks to Alex, Daniel, John and Scott for all feedback!
>
> v5:
>   * Aligned the implementation to ixgbe's XDP support: naming, favor
>     xchg instead of RCU semantics
>   * Support for XDP headroom (piggybacking on Alex' build_skb work)
>   * Added xdp tracepoints for exception states (as suggested by
>     Daniel)
>
> v4:
>   * Removed unused i40e_page_is_reserved function
>   * Prior running the XDP program, set the struct xdp_buff
>     data_hard_start member
>
> v3:
>   * Rebased patch set on Jeff's dev-queue branch
>   * MSI-X is no longer a prerequisite for XDP
>   * RCU locking for the XDP program and XDP_RX support is introduced
>     in the same patch
>   * Rx bytes is now bumped for XDP
>   * Removed pointer-to-pointer clunkiness
>   * Added comments to XDP preconditions in ndo_xdp
>   * When a non-EOF is received, log once, and drop the frame
>
> v2:
>   * Fixed kbuild error for PAGE_SIZE >= 8192.
>   * Renamed i40e_try_flip_rx_page to i40e_can_reuse_rx_page, which is
>     more in line to the other Intel Ethernet drivers (igb/fm10k).
>   * Validate xdp_adjust_head support in ndo_xdp/XDP_SETUP_PROG.
>
> Björn Töpel (2):
>   i40e: add XDP support for pass and drop actions
>   i40e: add support for XDP_TX action
>
>  drivers/net/ethernet/intel/i40e/i40e.h         |   8 +
>  drivers/net/ethernet/intel/i40e/i40e_ethtool.c |  57 +++++-
>  drivers/net/ethernet/intel/i40e/i40e_main.c    | 270 ++++++++++++++++++++++---
>  drivers/net/ethernet/intel/i40e/i40e_txrx.c    | 245 ++++++++++++++++++----
>  drivers/net/ethernet/intel/i40e/i40e_txrx.h    |  12 ++
>  5 files changed, 530 insertions(+), 62 deletions(-)
>
> --
> 2.11.0
>