Re: [PATCH net-next v5 0/8] net: macb: Add XDP support and page pool integration

[Nicolai Buchwitz <nb@tipi-net.de>]

On 13.3.2026 21:14, Paolo Valerio wrote:
> Tested on Raspberry Pi 5.
> With rpi5 is not possible to test the !RSC case. It would be nice
> to test it, as it potentially changes the offset of the first frame
> (hw shifts of NET_IP_ALIGN bytes on rx).
> 
> All the changes are intended for gem only.
> 
> The series consists of two main changes:
> 
> - Migration from netdev_alloc_skb() to page pool allocation model,
>   enabling skb recycling.
>   This also adds support for multi-descriptor frame reception,
>   removing the previous single-descriptor approach and avoiding
>   potentially large contiguous allocations for e.g. jumbo frames
>   with CONFIG_PAGE_SIZE_4KB.
> 
> - XDP support: Initial XDP implementation supporting major
>   verdicts (XDP_PASS, XDP_DROP, XDP_REDIRECT, XDP_TX) along with
>   the ndo_xdp_xmit function for packet redirection.
> 
> The driver now advertises NETDEV_XDP_ACT_BASIC, 
> NETDEV_XDP_ACT_REDIRECT,
> NETDEV_XDP_ACT_NDO_XMIT capabilities.
> 
> Some numbers (Andrew)
> =====================
> 
>   Initially reported numbers (Mbps, w/ drops on the rx side):
> 
>   
> https://lore.kernel.org/netdev/b0bff7a8-a004-4eb7-bf1d-2137182e59f9@lunn.ch/
> 
>   The numbers are essentially always the same with these two packet 
> sizes.
>   cpu0 doesn't appear overloaded. Not sure if there's a bottlneck at 
> the
>   interface level on my hw, but it seems cpu can handle more.
> 
>   Second test run (with no drops on the rx side):
> 
>   tx side: iperf3 sending UDP (256B) packets roughly at line rate (no 
> drops
>   on the receiving side).
> 
>   rx side:
>   - fixed cpu frequency
>   - single rxq on cpu 0
>   - iperf3 server on cpu 1-3
>   - average w/ mpstat 20s interval 20s count
> 
>   Baseline:
> 
>   Average:     CPU    %usr    %sys    %soft  %idle
>   Average:     all    3.64   15.05     0.87  80.42
>   Average:       0    0.05    0.15    10.23  89.56
>   Average:       1    4.20   16.98     0.00  78.81
>   Average:       2    3.67   15.53     0.00  80.79
>   Average:       3    4.07   16.83     0.00  79.09
> 
>   Page pool:
> 
>   Average:     CPU    %usr    %sys    %soft  %idle
>   Average:     all    3.76   14.91     0.74  80.57
>   Average:       0    0.05    0.12     7.81  91.98
>   Average:       1    3.88   15.53     0.00  80.58
>   Average:       2    4.35   17.23     0.00  78.42
>   Average:       3    4.22   16.63     0.00  79.11
> 
> 
> Previous versions
> =================
>   - v4: 
> https://lore.kernel.org/all/20260309144353.1213770-1-pvalerio@redhat.com/
>   - v3: 
> https://lore.kernel.org/netdev/20260302115232.1430640-1-pvalerio@redhat.com/
>   - v2: 
> https://lore.kernel.org/netdev/20260223182632.1681809-1-pvalerio@redhat.com/
>   - v1: 
> https://lore.kernel.org/netdev/20260115222531.313002-1-pvalerio@redhat.com/
>   - RFC v2: 
> https://lore.kernel.org/netdev/20251220235135.1078587-1-pvalerio@redhat.com/
>   - RFC v1: 
> https://lore.kernel.org/netdev/20251119135330.551835-1-pvalerio@redhat.com/
> 
> v4 -> v5
> ========
>   - Call missing macb_tx_lpi_wake() in macb_xdp_submit_frame() (Nicolai 
> Buchwitz)
>   - Restore old_prog (instead of NULL) in gem_xdp_setup() on 
> macb_open() failure
> 
> v3 -> v4
> ========
>   - Remove inline from gem_rx_data_len() (Paolo Abeni)
>   - Rate limit the netdev_{err,warn} in gem_rx() and gem_rx_refill() 
> with to avoid log
>     flooding under error conditions (Paolo Abeni)
>   - Fix missing newline in netdev_err() for MTU validation in 
> ndo_change_mtu (Paolo Abeni)
> 
> v2 -> v3
> ========
>   - Drop mog_init_rings() call in macb_open() reintroduced by mistake 
> while
>     rebasing (Jakub Kicinski)
>   - Move [5/8] down in the series (was [6/8])
>   - s/void/sk_buff for the skb parameter instead of updating the 
> documentation
>     because of its void type as this seems to better fit the scope of 
> the
>     patch (Jakub Kicinski)
>   - Move [6/8] down in the series (was [7/8])
>   - Remove misleading check on tx buff type (MACB_TYPE_SKB) to improve
>     clarity as all buffers are MACB_TYPE_SKB (Jakub Kicinski)
>   - Update commit message to better fit with the patch content
>   - Move [7/8] up in the series (was [5/8]) as XDP_TX has now been 
> folded in
>     this patch honoring the advertised features (Jakub Kicinski)
>   - explicitly return 0 in gem_create_page_pool() instead of err (Jakub 
> Kicinski)
>   - Revisit gem_xdp_valid_mtu() and related helpers lifting the 
> dependency with
>     bp->rx_headroom. It now always uses XDP_PACKET_HEADROOM (Jakub 
> Kicinski)
>   - Use xdp_return_frame() in workqueue context (Jakub Kicinski)
>   - macb_tx_error_task() now locks tx_ptr_lock (Jakub Kicinski)
>   - Remove redundant check for clarity (Jakub Kicinski)
>   - Take into account head adjustment in macb_xdp_submit_frame() while 
> syncing
>     in the XDP_TX path and for the XDP_PASS codepath.
>   - Some helpers renaming for clarity
>   - Do not to leave bp->prog pointer in gem_setup_xdp dangling on 
> failure
>   - Revisit rx_buffer_size to better handle the RSC/non-RSC case
>   - Return error in ndo_xdp_xmit if the interface is not operating
>   - Handle packets count in macb_tx_complete() for the xdp case
> 
> v1 -> v2
> ========
>   - Fix potential NULL pointer dereference in gem_rx() when receiving 
> an
>     unexpected non-starting frame without an active skb (Jakub 
> Kicinski).
>   - Make sure to release both the pending skb and the current buff in 
> fragment
>     overflow error path (Jakub Kicinski).
>   - Reuse the existing page pool instead of creating a new one in HRESP 
> error
>     recovery (Jakub Kicinski).
>   - s/page/buffer/ in netdev_err() print out (Théo Lebrun)
>   - Ensure pending queue->skb is released in gem_free_rx_buffers() 
> (Théo Lebrun).
>   - Use NET_SKB_PAD for the headroom when no xdp program is attached
>   - Add netdev to pp_params (for pp stats)
>   - Move ip_summed handling when RX_EOF is set as RX_SUM may not be 
> reliable when
>     RX_EOF is not set.
>   - Sync for device no longer relies on PP_FLAG_DMA_SYNC_DEV reducing 
> the
>     sync size.
>   - Introduce rx_ip_align field in struct macb
>   - Fix incorrect DMA direction for page pool when attaching XDP 
> program by
>     restoring close/open sequence as in a previous version (Jakub 
> Kicinski).
>   - Rename release_buff() to macb_tx_release_buff() for clarity (Théo 
> Lebrun).
>   - Do not unmap dma for XDP_TX buffers by ensuring tx_buff->mapping is 
> zero for
>     MACB_TYPE_XDP_TX buff_type (Jakub Kicinski).
>   - Always return memory to pool on xdp_convert_buff_to_frame()
>     failure (Jakub Kicinski).
>   - Remove skip_xdp label (Théo, this is a bit different from your 
> suggestion,
>     let me know if you have a strong preference here)
> 
> RFC v2 -> v1
> ============
>   - Removed bp->macbgem_ops.mog_init_rings(bp) call from macb_open()
>   - Fixed includes (remove unneeded, moved one from header to 
> macb_main.c)
>   - Reverse xmas tree ordering (gem_rx, gem_rx_refill)
>   - print_hex_dump_debug() instead of print_hex_dump()
>   - Replaced rx frame length check with MACB_BIT(RX_EOF) for data_len
>     calculation
>   - Removed NET_IP_ALIGN handling in rx buffer size calculation
>   - Updated debug format string to include rx_headroom and total size
>   - Changed types to unsigned int in helper functions and variable
>   - Removed unneeded line break
> 
> RFC v1 -> RFC v2
> ================
>   - Squashed 1/6 and 2/6
>   - Reworked rx_buffer_size computation. It no longer takes into
>     accounts extra room.
>   - A bunch of renaming (rx_offset => rx_headroom, removed 
> MACB_MAX_PAD,
>     MACB_PP_HEADROOM => XDP_PACKET_HEADROOM, data => ptr, xdp_q => 
> xdp_rxq,
>     macb_xdp() => gem_xdp(), macb_xdp_xmit() => gem_xdp_xmit())
>   - Deduplicated buffer size computation in gem_xdp_valid_mtu()
>     and gem_xdp_setup()
>   - gem_xdp_setup() no longer close()/open()
>   - Renaming from rx_skbuff to rx_buff is now got split in a separate 
> commit
>   - Open-coded gem_page_pool_get_buff()
>   - Added missing rx_buff re-initialization in the error path during rx
>   - Page pool creation failure now fails the device open
>   - Moved xdp buff preparation inside gem_xdp_run()
>   - Added missing rcu_access_pointer()
>   - Turned return value in -EOPNOTSUPP for macb_xdp() on failure
>   - moved tx_skb to tx_buff renaming to a separate commit
>   - Removed some unneeded code and set MACB_TYPE_SKB for 
> lp->rm9200_txq[desc].type as well
>   - Replaced !!addr with a dedicated bool in macb_xdp_submit_frame()
> 
> Paolo Valerio (7):
>   net: macb: rename rx_skbuff into rx_buff
>   net: macb: Add page pool support handle multi-descriptor frame rx
>   net: macb: use the current queue number for stats
>   net: macb: make macb_tx_skb generic
>   net: macb: generalize tx buffer handling
>   net: macb: add XDP support for gem
>   net: macb: introduce ndo_xdp_xmit support
> 
> Théo Lebrun (1):
>   net: macb: move Rx buffers alloc from link up to open
> 
>  drivers/net/ethernet/cadence/Kconfig     |   1 +
>  drivers/net/ethernet/cadence/macb.h      |  43 +-
>  drivers/net/ethernet/cadence/macb_main.c | 951 ++++++++++++++++++-----
>  3 files changed, 770 insertions(+), 225 deletions(-)

Tested on Raspberry Pi 5:
- XDP attach/detach
- XDP_PASS (no throughput regression at 938 Mbit/s)
- XDP_DROP
- XDP_TX

The driver survives link flaps and rapid program swaps under iperf3 load 
without errors.
MTU validation correctly rejects XDP when frames would span multiple 
pages.

Tested-by: Nicolai Buchwitz <nb@tipi-net.de>