This is version 6 of the AF_XDP zero-copy support for igb. Since Sriram's duties changed I am sending this instead. Additionally, I've tested this on real hardware, Intel i210 [1].
Changes since v5: - Rebase to 6.11 - Fix set-but-unused variable warnings - Split first patches (Maciej) - Add READ/WRITE_ONCE() for xsk_pool and xdp_prog (Maciej) - Add synchronize_net() (Maciej) - Remove IGB_RING_FLAG_AF_XDP_ZC (Maciej) - Add NETDEV_XDP_ACT_XSK_ZEROCOPY to last patch (Maciej) - Update Rx ntc handling (Maciej) - Move stats update and xdp finalize to common functions (Maciej) - "Likelyfy" XDP_REDIRECT case (Maciej) - Check Tx disabled and carrier in igb_xmit_zc() (Maciej) - RCT (Maciej) - Link to v5: https://lore.kernel.org/r/20240711-b4-igb_zero_copy-v5-0-f3f455113...@linutronix.de Changes since v4: - Rebase to v6.10 - Fix issue reported by kernel test robot - Provide napi_id for xdp_rxq_info_reg() so that busy polling works - Set olinfo_status in igb_xmit_zc() so that frames are transmitted Link to v4: https://lore.kernel.org/intel-wired-lan/20230804084051.14194-1-sriram.yagnara...@est.tech/ [1] - https://github.com/Linutronix/TSN-Testbench/tree/main/tests/busypolling_i210 Original cover letter: The first couple of patches adds helper funcctions to prepare for AF_XDP zero-copy support which comes in the last couple of patches, one each for Rx and TX paths. As mentioned in v1 patchset [0], I don't have access to an actual IGB device to provide correct performance numbers. I have used Intel 82576EB emulator in QEMU [1] to test the changes to IGB driver. The tests use one isolated vCPU for RX/TX and one isolated vCPU for the xdp-sock application [2]. Hope these measurements provide at the least some indication on the increase in performance when using ZC, especially in the TX path. It would be awesome if someone with a real IGB NIC can test the patch. AF_XDP performance using 64 byte packets in Kpps. Benchmark: XDP-SKB XDP-DRV XDP-DRV(ZC) rxdrop 220 235 350 txpush 1.000 1.000 410 l2fwd 1.000 1.000 200 AF_XDP performance using 1500 byte packets in Kpps. Benchmark: XDP-SKB XDP-DRV XDP-DRV(ZC) rxdrop 200 210 310 txpush 1.000 1.000 410 l2fwd 0.900 1.000 160 [0]: https://lore.kernel.org/intel-wired-lan/20230704095915.9750-1-sriram.yagnara...@est.tech/ [1]: https://www.qemu.org/docs/master/system/devices/igb.html [2]: https://github.com/xdp-project/bpf-examples/tree/master/AF_XDP-example v3->v4: - NULL check buffer_info in igb_dump before dereferencing (Simon Horman) v2->v3: - Avoid TX unit hang when using AF_XDP zero-copy by setting time_stamp on the tx_buffer_info - Fix uninitialized nb_buffs (Simon Horman) v1->v2: - Use batch XSK APIs (Maciej Fijalkowski) - Follow reverse xmas tree convention and remove the ternary operator use (Simon Horman) --- Sriram Yagnaraman (6): igb: Always call igb_xdp_ring_update_tail() under Tx lock igb: Remove static qualifiers igb: Introduce igb_xdp_is_enabled() igb: Introduce XSK data structures and helpers igb: Add AF_XDP zero-copy Rx support igb: Add AF_XDP zero-copy Tx support drivers/net/ethernet/intel/igb/Makefile | 2 +- drivers/net/ethernet/intel/igb/igb.h | 36 +- drivers/net/ethernet/intel/igb/igb_main.c | 232 ++++++++---- drivers/net/ethernet/intel/igb/igb_xsk.c | 561 ++++++++++++++++++++++++++++++ 4 files changed, 770 insertions(+), 61 deletions(-) --- base-commit: e7d731326ef0622f103e5ed47d3405f71cdcd7f6 change-id: 20240711-b4-igb_zero_copy-bb70a31ecb0f Best regards, -- Kurt Kanzenbach <k...@linutronix.de>
