[RFC v1] net/af_packet: add io_uring TX acceleration

[Stephen Hemminger]

On Wed,  4 Feb 2026 08:12:02 -0800
scott.k.mitch1 at gmail.com wrote:

> From: Scott <scott.k.mitch1 at gmail.com>
> 
> When user space application does sendto() to notify the kernel
> there is data to process, the kernel synchronously processes
> pending frames and the CPU cost is attributed to the application.
> The kernel processing may also include kernel RX (netlink, ovs)
> and packet manipulation. This can negatively impact performance
> by limiting CPU to each dpdk network thread.
> 
> io_uring offers a mode (SQPOLL) to offload this processing
> to a dedicated kernel thread. This makes the kernel processing
> asynchronous, which can improve throughput by 30%+. The trade-off
> is SQPOLL threads consume additional CPU but this follows the
> same principle as virtio backends, and enables use cases that
> can't adopt virtio (e.g. containerized workloads).
> 
> This patch adds optional io_uring support for TX kick operations
> to replace synchronous sendto() syscalls. When enabled, uses
> SQPOLL mode with kernel-side submission polling for reduced
> syscall overhead.
> 
> Features:
> - Optional liburing dependency (version >= 2.4 required)
> - SQPOLL mode with configurable idle timeout
> - CPU affinity for SQPOLL threads (IORING_SETUP_SQ_AFF)
> - Shared workqueue support (IORING_SETUP_ATTACH_WQ)
> - Registered file descriptors for IOSQE_FIXED_FILE optimization
> - Single issuer and cooperative taskrun optimizations when available
> 
> New devargs:
> - tx_io_uring_enabled: Enable io_uring TX (default: 0)
> - tx_io_uring_sq_size: Submission queue size (default: 1024)
> - tx_io_uring_sq_thread_idle: SQPOLL idle timeout in ms (default: 500)
> - tx_io_uring_q_cpu: CPU affinity for SQPOLL threads (repeatable)
> - tx_io_uring_wq_num: Number of shared workqueues (default: 0)
> 
> The TX path updates the ring tail pointer directly instead of calling
> io_uring_submit() to avoid the io_uring_enter syscall, relying on
> SQPOLL to pick up submissions. If the SQPOLL thread sleeps, we wake
> it with IORING_ENTER_SQ_WAKEUP.
> 
> Falls back gracefully to sendto() when:
> - liburing is not available at compile time
> - SQPOLL is not supported by the kernel
> - io_uring initialization fails for any reason
> 
> Signed-off-by: Scott <scott.k.mitch1 at gmail.com>
> ---
> Depends-on: series-37248 ("af_packet correctness, performance, cksum")
> 
>  doc/guides/nics/af_packet.rst             |  16 +-
>  doc/guides/rel_notes/release_26_03.rst    |   1 +
>  drivers/net/af_packet/meson.build         |  14 +
>  drivers/net/af_packet/rte_eth_af_packet.c | 420 +++++++++++++++++++++-
>  4 files changed, 435 insertions(+), 16 deletions(-)
> 
> diff --git a/doc/guides/nics/af_packet.rst b/doc/guides/nics/af_pac

I would rather it be a different driver than af_packet.
See my RFC patch for using io_uring