Re: [net-next v18 5/7] net: mtip: Add mtip_switch_{rx|tx} functions to the L2 switch driver

[Łukasz Majewski]

Hi Jakub,

> On Wed, 13 Aug 2025 09:07:53 +0200 Lukasz Majewski wrote:
> > +		page = fep->page[bdp - fep->rx_bd_base];
> > +		/* Process the incoming frame */
> > +		pkt_len = bdp->cbd_datlen;
> > +
> > +		dma_sync_single_for_cpu(&fep->pdev->dev,
> > bdp->cbd_bufaddr,
> > +					pkt_len, DMA_FROM_DEVICE);
> > +		net_prefetch(page_address(page));
> > +		data = page_address(page);
> > +
> > +		if (fep->quirks & FEC_QUIRK_SWAP_FRAME)
> > +			swap_buffer(data, pkt_len);
> > +
> > +		eth_hdr = (struct ethhdr *)data;
> > +		mtip_atable_get_entry_port_number(fep,
> > eth_hdr->h_source,
> > +						  &rx_port);
> > +		if (rx_port == MTIP_PORT_FORWARDING_INIT)
> > +			mtip_atable_dynamicms_learn_migration(fep,
> > +
> > mtip_get_time(),
> > +
> > eth_hdr->h_source,
> > +
> > &rx_port); +
> > +		if ((rx_port == 1 || rx_port == 2) &&
> > fep->ndev[rx_port - 1])
> > +			pndev = fep->ndev[rx_port - 1];
> > +		else
> > +			pndev = dev;
> > +
> > +		*port = rx_port;
> > +
> > +		/* This does 16 byte alignment, exactly what we
> > need.
> > +		 * The packet length includes FCS, but we don't
> > want to
> > +		 * include that when passing upstream as it messes
> > up
> > +		 * bridging applications.
> > +		 */
> > +		skb = netdev_alloc_skb(pndev, pkt_len +
> > NET_IP_ALIGN);
> > +		if (unlikely(!skb)) {
> > +			dev_dbg(&fep->pdev->dev,
> > +				"%s: Memory squeeze, dropping
> > packet.\n",
> > +				pndev->name);
> > +			page_pool_recycle_direct(fep->page_pool,
> > page);
> > +			pndev->stats.rx_dropped++;
> > +			return -ENOMEM;
> > +		}
> > +
> > +		skb_reserve(skb, NET_IP_ALIGN);
> > +		skb_put(skb, pkt_len);      /* Make room */
> > +		skb_copy_to_linear_data(skb, data, pkt_len);
> > +		skb->protocol = eth_type_trans(skb, pndev);
> > +		skb->offload_fwd_mark = fep->br_offload;
> > +		napi_gro_receive(&fep->napi, skb);  
> 
> The rx buffer circulation is very odd.

The fec_main.c uses page_pool_alloc_pages() to allocate RX page from
the pool.

At the RX function the __build_skb(data, ...) is called to create skb.

Last step with the RX function is to call skb_mark_for_recycle(skb),
which sets skb->pp_recycle = 1.

And yes, in the MTIP I do copy the data to the newly created skb in RX
function (anyway, I need to swap bytes in the buffer). 

It seems like extra copy is performed in the RX function.

> You seem to pre-allocate
> buffers for the full ring from a page_pool. And then copy the data
> out of those pages.

Yes, correct.

> The normal process is that after packet is
> received a new page is allocated to give to HW, and old is attached
> to an skb, and sent up the stack.

Ok.

> 
> Also you are releasing the page to be recycled without clearing it
> from the ring. I think you'd free it again on shutdown, so it's a
> double-free.

No, the page is persistent. It will be removed when the driver is
closed and memory for pages and descriptors is released.

-- 
Best regards,

Łukasz Majewski