Add XDP support (w/o XSk for now) to the idpf driver using the libeth_xdp
sublib. All possible verdicts, .ndo_xdp_xmit(), multi-buffer etc. are here.
In general, nothing outstanding comparing to ice, except performance --
let's say, up to 2x for .ndo_xdp_xmit() on certain platforms and
scenarios.
idpf doesn't support VLAN Rx offload, so only the hash hint is
available for now.
Patches 1-7 are prereqs, without which XDP would either not work at all or
work slower/worse/...
Alexander Lobakin (9):
xdp, libeth: make the xdp_init_buff() micro-optimization generic
idpf: fix Rx descriptor ready check barrier in splitq
idpf: use a saner limit for default number of queues to allocate
idpf: link NAPIs to queues
idpf: add support for nointerrupt queues
idpf: use generic functions to build xdp_buff and skb
idpf: add support for XDP on Rx
idpf: add support for .ndo_xdp_xmit()
idpf: add XDP RSS hash hint
Michal Kubiak (4):
idpf: add 4-byte completion descriptor definition
idpf: remove SW marker handling from NAPI
idpf: prepare structures to support XDP
idpf: implement XDP_SETUP_PROG in ndo_bpf for splitq
drivers/net/ethernet/intel/idpf/Kconfig | 2 +-
drivers/net/ethernet/intel/idpf/Makefile | 2 +
drivers/net/ethernet/intel/idpf/idpf.h | 31 +-
.../net/ethernet/intel/idpf/idpf_lan_txrx.h | 6 +-
drivers/net/ethernet/intel/idpf/idpf_txrx.h | 140 ++++--
.../net/ethernet/intel/idpf/idpf_virtchnl.h | 1 -
drivers/net/ethernet/intel/idpf/xdp.h | 172 +++++++
include/net/libeth/xdp.h | 11 +-
include/net/xdp.h | 28 +-
drivers/net/ethernet/intel/idpf/idpf_dev.c | 11 +-
drivers/net/ethernet/intel/idpf/idpf_lib.c | 67 ++-
drivers/net/ethernet/intel/idpf/idpf_main.c | 1 +
.../ethernet/intel/idpf/idpf_singleq_txrx.c | 110 ++---
drivers/net/ethernet/intel/idpf/idpf_txrx.c | 424 ++++++++--------
drivers/net/ethernet/intel/idpf/idpf_vf_dev.c | 11 +-
.../net/ethernet/intel/idpf/idpf_virtchnl.c | 173 ++++---
drivers/net/ethernet/intel/idpf/xdp.c | 454 ++++++++++++++++++
17 files changed, 1217 insertions(+), 427 deletions(-)
create mode 100644 drivers/net/ethernet/intel/idpf/xdp.h
create mode 100644 drivers/net/ethernet/intel/idpf/xdp.c
---
From v4[0]:
* 08/13: fix rare crashes when the number of XDPQSs < nr_cpu_ids -- merge
artefacts after rebasing on top of Joshua's recent changes;
* 09/13: fix "suspicious RCU usage" during rmmod when lockdep is enabled.
From v3[1]:
* 01/13: make the xdp_init_buff() micro-opt generic, include some
bloat-o-meter and perf diffs (Simon, Kees);
* 08/13: don't include XDPSQs in Ethtool's 'other_count' (Ethtool
channels are interrupts!) (Jakub);
* 11/13:
* finalize XDPSQs a bit earlier on Rx;
* show some bloat-o-meter and performance diffs for
__LIBETH_WORD_ACCESS (Jakub).
From v2[2]:
* rebase on top of [3] to resolve conflicts in Tony's tree;
* 02:
* harmonize maximum number of queues to not create more Tx queues than
completion queues or more Rx queues than buffer queues / 2;
* fix VC timeouts on certain steppings as there processing a lot of queues
can take more time than the minimum timeout of 2 seconds;
* 03: fix RTNL assertion fail on PCI reset.
From v1[4]:
* drop the libeth_xdp part (submitted separately and accepted);
* fix some typos and kdocs (Jakub, Maciej);
* pick a couple RBs (Maciej);
* 03: create a convenience helper (Maciej), fix rtnl assertion fail;
* 04: since XDP uses its own queue cleaning routines, don't add 4-byte
completion support to the skb code;
* 05: don't use old weird logic with negative descriptor index (Maciej);
* 06: fix invalid interrupt vector counting in certain cases;
* 07: fix cleanup timer is fired after the queue buffers are already freed;
* 08: fix XDP program removal in corner cases such as PCI reset or
remove request when there's no active prog (from netdev_unregister());
* 10: fix rare queue stuck -- HW requires to always have at least one free Tx
descriptor on the queue, otherwise it thinks the queue is empty and
there's nothing to send (true Intel HW veteran bug).
Testing hints: basic Rx and Tx (TCP, UDP, VLAN, HW GRO on/off, trafficgen
stress tests, performance comparison); xdp-tools with all possible actions
(xdp-bench for PASS, DROP, TX, REDIRECT to cpumap, devmap (inc self-redirect);
xdp-trafficgen to double-check XDP xmit). Would be nice to see a perf
comparison against ice (in percent) (idpf must be plugged into a PCIe 4+).
[0] https://lore.kernel.org/intel-wired-lan/20250811161044.32329-1-aleksander.lobakin@intel.com
[1] https://lore.kernel.org/intel-wired-lan/20250730160717.28976-1-aleksander.lobakin@intel.com
[2] https://lore.kernel.org/intel-wired-lan/20250624164515.2663137-1-aleksander.lobakin@intel.com
[3] https://lore.kernel.org/intel-wired-lan/20250725184223.4084821-1-joshua.a.hay@intel.com
[4] https://lore.kernel.org/intel-wired-lan/20250305162132.1106080-1-aleksander.lobakin@intel.com
--
2.51.0