On 23 Jan 2026 at 12:04:55 AM, Andrew Lunn [off-list ref] wrote:

On Thu, Jan 22, 2026 at 11:24:05PM +0100, Paolo Valerio wrote:

quoted

On 16 Jan 2026 at 06:16:16 PM, Andrew Lunn [off-list ref] wrote:

quoted

On Thu, Jan 15, 2026 at 11:25:26PM +0100, Paolo Valerio wrote:

quoted

Use the page pool allocator for the data buffers and enable skb recycling
support, instead of relying on netdev_alloc_skb allocating the entire skb
during the refill.

Do you have any benchmark numbers for this change? Often swapping to
page pool improves the performance of the driver, and i use it as a
selling point for doing the conversion, independent of XDP.

I finally got the chance to get my hands on the board.

On the rpi5 I simply run xdp-bench in skb-mode to drop and collect the
stats.

Page size is 4k and stats include the driver consuming a full page
setting mtu such that rx_buffer_size + overhead exceed half page and the
other way around for 2 fragments.

             |    64     |    128    |
  baseline   |  533,158  |  531,618  |
  pp page    |  530,929  |  529,682  |
  pp 2 frags |  530,781  |  529,116  |

I was more interested in plain networking, not XDP. Does it perform
better with page pool? You at least need to show it is not worse, you
need to avoid performance regressions.

I retested with iperf3. The target has a single rx queue with iperf3
running with no cpu affinity set.

|              |  64 | 128 |
| baseline     | 273 | 545 |
| pp (page)    | 273 | 544 |
| pp (2 frags) | 272 | 544 |

     Andrew