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

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

[Karl Mehltretter]

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.

Full test details and logs are available if useful.

Thanks,
Karl

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

[Greg Kroah-Hartman]

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?

Also, there is no context here in the email, please always include that
when replying.

thanks,

greg k-h

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

[Nicolai Buchwitz]

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

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

[Nicolai Buchwitz]

Hi Greg, hi Karl

On 24.9.2026 09:28, Greg Kroah-Hartman wrote:
> On Thu, Sep 24, 2026 at 08:58:39AM +0200, Karl Mehltretter wrote:

> [...]

>> I would appreciate the BCMGENET maintainers' view on whether this 
>> page-pool
>> conversion is appropriate for 6.18.y.

I would rather have the page_pool conversion in 6.18.y. More fixes will
land on top of it, and carrying it is less work than adapting each of
them. The fix for the regression applies on top of it unchanged.

>> 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?

The fix is on netdev, "net: bcmgenet: allocate RX buffers as page
fragments", with a Fixes tag for 7bc054c2d4ed:

   https://lore.kernel.org/netdev/20260924101922.2675127-1-nb@tipi-net.de

I was able to reproduce Karl's numbers on a CM4 with a 64 KiB page 
kernel
and fortunately the fix was straightforward.

> [...]

Regards
Nicolai