linux/ena: Free copybreak page if NUMA is incorrect

RX page allocation mechanism allocates pages on the NUMA node associated
with the CPU which serves the queue's napi routine.
Specifically, in LPC case, we check the allocated cache pages per
RX refill routine, and re-allocate them in case they are located in the
wrong NUMA node.

These mechanisms were not applied to copybreak pages, as their content
is being copied and the pages are being reused, without checking
whether they are allocated on the correct NUMA node.
This patch makes sure that an RX page is re-allocated if it's on the
wrong NUMA for packets on the copybreak path.

Fixes: 798070a ("First commit of Amazon Elastic Network Adapter (ENA) Linux driver")
Signed-off-by: Shahar Itzko <itzko@amazon.com>

kernel/linux/ena/ena_netdev.c

@@ -1198,6 +1198,11 @@ static struct sk_buff *ena_rx_skb(struct ena_ring *rx_ring,
 		rx_ring->free_ids[*next_to_clean] = req_id;
 		*next_to_clean = ENA_RX_RING_IDX_ADD(*next_to_clean, descs,
 						     rx_ring->ring_size);
+
+		/* Don't reuse the RX page if we're on the wrong NUMA */
+		if (page_to_nid(rx_info->page) != numa_mem_id())
+			ena_free_rx_page(rx_ring, rx_info);
+
 		return skb;
 	}