[iwl-next,00/12] idpf: XDP chapter I: convert Rx to libeth

Message

Alexander Lobakin May 28, 2024, 1:48 p.m. UTC

XDP for idpf is currently 5 chapters:
* convert Rx to libeth (this);
* convert Tx and stats to libeth;
* generic XDP and XSk code changes, libeth_xdp;
* actual XDP for idpf via libeth_xdp;
* XSk for idpf (^).

Part I does the following:
* splits &idpf_queue into 4 (RQ, SQ, FQ, CQ) and puts them on a diet;
* ensures optimal cacheline placement, strictly asserts CL sizes;
* moves currently unused/dead singleq mode out of line;
* reuses libeth's Rx ptype definitions and helpers;
* uses libeth's Rx buffer management for both header and payload;
* eliminates memcpy()s and coherent DMA uses on hotpath, uses
  napi_build_skb() instead of in-place short skb allocation.

Most idpf patches, except for the queue split, removes more lines
than adds.
Expect far better memory utilization and +5-8% on Rx depending on
the case (+17% on skb XDP_DROP :>).

Alexander Lobakin (12):
  libeth: add cacheline / struct alignment helpers
  idpf: stop using macros for accessing queue descriptors
  idpf: split &idpf_queue into 4 strictly-typed queue structures
  idpf: avoid bloating &idpf_q_vector with big %NR_CPUS
  idpf: strictly assert cachelines of queue and queue vector structures
  idpf: merge singleq and splitq &net_device_ops
  idpf: compile singleq code only under default-n CONFIG_IDPF_SINGLEQ
  idpf: reuse libeth's definitions of parsed ptype structures
  idpf: remove legacy Page Pool Ethtool stats
  libeth: support different types of buffers for Rx
  idpf: convert header split mode to libeth + napi_build_skb()
  idpf: use libeth Rx buffer management for payload buffer

 drivers/net/ethernet/intel/Kconfig            |   13 +-
 drivers/net/ethernet/intel/idpf/Kconfig       |   26 +
 drivers/net/ethernet/intel/idpf/Makefile      |    3 +-
 scripts/kernel-doc                            |    1 +
 drivers/net/ethernet/intel/idpf/idpf.h        |   11 +-
 .../net/ethernet/intel/idpf/idpf_lan_txrx.h   |    2 +
 drivers/net/ethernet/intel/idpf/idpf_txrx.h   |  800 +++++-----
 include/net/libeth/cache.h                    |  100 ++
 include/net/libeth/rx.h                       |   19 +
 .../net/ethernet/intel/idpf/idpf_ethtool.c    |  152 +-
 drivers/net/ethernet/intel/idpf/idpf_lib.c    |   88 +-
 drivers/net/ethernet/intel/idpf/idpf_main.c   |    1 +
 .../ethernet/intel/idpf/idpf_singleq_txrx.c   |  311 ++--
 drivers/net/ethernet/intel/idpf/idpf_txrx.c   | 1410 +++++++++--------
 .../net/ethernet/intel/idpf/idpf_virtchnl.c   |  178 ++-
 drivers/net/ethernet/intel/libeth/rx.c        |  132 +-
 16 files changed, 1824 insertions(+), 1423 deletions(-)
 create mode 100644 drivers/net/ethernet/intel/idpf/Kconfig
 create mode 100644 include/net/libeth/cache.h

---
Applies on top of "idpf: don't enable NAPI and interrupts prior to
allocating Rx buffers" from the -net tree, otherwise may work unstable.

From RFC[0]:
*  *: add kdocs where needed and fix the existing ones to build cleanly;
      fix minor checkpatch and codespell warnings;
      add RBs from Przemek;
* 01: fix kdoc script to understand new libeth_cacheline_group() macro;
      add an additional assert for queue struct alignment;
* 02: pick RB from Mina;
* 06: make idpf_chk_linearize() static as it's now used only in one file;
* 07: rephrase the commitmsg: HW supports it, but never wants;
* 08: fix crashes on some configurations (Mina);
* 11: constify header buffer pointer in idpf_rx_hsplit_wa().

Testing hints: basic Rx regression tests (+ perf and memory usage
before/after if needed).

Note that I'm on a vacation starting 30th May and return 10th of June.
I may sometimes reply, but no new versions in the meantime =\

[0] https://lore.kernel.org/netdev/[e-mail]