On Thu, 12 Jun 2025 23:02:30 +0200 Lorenzo Bianconi wrote:

quoted

quoted

I'm not Eric but FWIW 256B is not going to help much. It's best to keep
the len / truesize ratio above 50%, so with 32k buffers we're talking
about copying multiple frames.  

what I mean here is reallocate the skb if the true size is small (e.g. below
256B) in order to avoid consuming the high order page from the page_pool. Maybe
we can avoid it if reducing the page order to 2 for LRO queues provide
comparable results.

Hm, truesize is the buffer size, right? If the driver allocated n bytes
of memory for packets it sent up the stack, the truesizes of the skbs
it generated must add up to approximately n bytes.

With 'truesize' I am referring to the real data size contained in the x-order
page returned by the hw. If this size is small, I was thinking to just allocate
a skb for it, copy the data from the x-order page into it and re-insert the
x-order page into the page_pool running page_pool_put_full_page().
Let me do some tests with order-2 page to see if the GRO can compensate the
reduced page size.

Regards,
Lorenzo

So if the HW places one aggregation session per buffer, and the buffer
is 32kB -- to avoid mem use ratio < 25% you'd need to copy all sessions
smaller than 8kB?

If I'm not making sense - just ignore, I haven't looked at the rest of
the driver :)