Re: [PATCH 6.18 129/398] net: bcmgenet: convert RX path to page_pool

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

Hi Greg

On 24.9.2026 09:28, Greg Kroah-Hartman wrote:
> On Thu, Sep 24, 2026 at 08:58:39AM +0200, Karl Mehltretter wrote:
>> I have a small objection to taking patch 129 into 6.18.y only as a
>> dependency for patch 130.
>> 
>> My LLM agent helped me running these tests.
>> 
>> On a 64 KiB arm64 kernel in a custom QEMU Pi 400 model with functional
>> GENET v5 DMA, patch 129 increased the socket receive-memory charge per
>> 512-byte UDP datagram from 2,752 to 65,792 bytes.
>> 
>> With the same 512 KiB socket receive-buffer limit, this reduced the 
>> number
>> of queued datagrams from 190 to 7 and increased socket receive-buffer 
>> drops
>> from 810 to 993 per 1,000 datagrams. I obtained the same result in 
>> five
>> bursts.
>> 
>> The page-pool conversion otherwise worked in this test, and physical 
>> Pi 400
>> testing with 4 KiB pages found no correctness failure. However, patch 
>> 130
>> was also verified on the Pi 400 to work without patch 129 after a 
>> small
>> contextual adaptation.
>> 
>> I would appreciate the BCMGENET maintainers' view on whether this 
>> page-pool
>> conversion is appropriate for 6.18.y.
>> 
>> Please consider omitting patch 129 from 6.18.y and applying that 
>> adapted
>> version of patch 130 directly.
> 
> But when you update to a newer release, that memory increase will
> happen, right?  So why not fix the root problem upstream first?

Agree. I will have a look at Karl's regression report and probably sent
patches via netdev with cc stable.

> [...]

Regards
Nicolai