On 2/16/2023 6:00 AM, Tirthendu Sarkar wrote:
This patchset adds multi-buffer support for XDP. Tx side already has
support for multi-buffer. This patchset focuses on Rx side. The last
patch contains actual multi-buffer changes while the previous ones are
preparatory patches.
On receiving the first buffer of a packet, xdp_buff is built and its
subsequent buffers are added to it as frags. While 'next_to_clean' keeps
pointing to the first descriptor, the newly introduced 'next_to_process'
keeps track of every descriptor for the packet.
On receiving EOP buffer the XDP program is called and appropriate action
is taken (building skb for XDP_PASS, reusing page for XDP_DROP, adjusting
page offsets for XDP_{REDIRECT,TX}).
The patchset also streamlines page offset adjustments for buffer reuse
to make it easier to post process the rx_buffers after running XDP prog.
With this patchset there does not seem to be any performance degradation
for XDP_PASS and some improvement (~1% for XDP_TX, ~5% for XDP_DROP) when
measured using xdp_rxq_info program from samples/bpf/ for 64B packets.
Changelog:
v4 -> v5:
- Change s/size/truesize [Tony]
- Rebased on top of commit 9dd6e53ef63d ("i40e: check vsi type before
setting xdp_features flag") [Lorenzo]
- Changed size of on stack variable to u32 from u16.
Hi Tirthendu,
Did you move over to next-queue/dev-queue because this series still
isn't applying.
Also, I'm not seeing the truesize change on patch 4 as there are still
issues being reported on it.
Thanks,
Tony
v3 -> v4:
- Added non-linear XDP buffer support to xdp_features. [Maciej]
- Removed double space. [Maciej]
v2 -> v3:
- Fixed buffer cleanup for single buffer packets on skb alloc
failure.
- Better naming of cleanup function.
- Stop incrementing nr_frags for overflowing packets.
v1 -> v2:
- Instead of building xdp_buff on eop now it is built incrementally.
- xdp_buff is now added to i40e_ring struct for preserving across
napi calls. [Alexander Duyck]
- Post XDP program rx_buffer processing has been simplified.
- Rx buffer allocation pull out is reverted to avoid performance
issues for smaller ring sizes and now done when at least half of
the ring has been cleaned. With v1 there was ~75% drop for
XDP_PASS with the smallest ring size of 64 which is mitigated by
v2 [Alexander Duyck]
- Instead of retrying skb allocation on previous failure now the
packet is dropped. [Maciej]
- Simplified page offset adjustments by using xdp->frame_sz instead
of recalculating truesize. [Maciej]
- Change i40e_trace() to use xdp instead of skb [Maciej]
- Reserve tailroom for legacy-rx [Maciej]
- Centralize max frame size calculation
Tirthendu Sarkar (8):
i40e: consolidate maximum frame size calculation for vsi
i40e: change Rx buffer size for legacy-rx to support XDP multi-buffer
i40e: add pre-xdp page_count in rx_buffer
i40e: Change size to truesize when using i40e_rx_buffer_flip()
i40e: use frame_sz instead of recalculating truesize for building skb
i40e: introduce next_to_process to i40e_ring
i40e: add xdp_buff to i40e_ring struct
i40e: add support for XDP multi-buffer Rx
drivers/net/ethernet/intel/i40e/i40e_main.c | 78 ++--
drivers/net/ethernet/intel/i40e/i40e_trace.h | 20 +-
drivers/net/ethernet/intel/i40e/i40e_txrx.c | 420 +++++++++++--------
drivers/net/ethernet/intel/i40e/i40e_txrx.h | 21 +-
4 files changed, 307 insertions(+), 232 deletions(-)
