ena_rx_skb() selects the copybreak path based on the first RX
descriptor's length without checking the number of descriptors. When
a packet spans multiple descriptors, this copies only the first
buffer into the skb and fails to recycle the remaining request IDs,
while RX cleanup advances by the full descriptor count.
Restrict copybreak to single-descriptor packets and apply the same
restriction to the rx_copybreak_pkt statistic. Multi-descriptor
packets then use the existing fragment-processing path.
Discovered by comparing the in-tree ENA driver with Amazon's
standalone ENA driver.
Compiled with W=1 under x86_64 defconfig.
Fixes: 1738cd3ed342 ("net: ena: Add a driver for Amazon Elastic Network Adapters (ENA)")
Signed-off-by: Avi Weiss <redacted>
---
Already fixed in Amazon's standalone ENA driver:
https://github.com/amzn/amzn-drivers/commit/8c47df628bbc6623bd9a419f261ccc46abbd36de
drivers/net/ethernet/amazon/ena/ena_netdev.c | 5 +++--
1 file changed, 3 insertions(+), 2 deletions(-)
diff --git a/drivers/net/ethernet/amazon/ena/ena_netdev.c b/drivers/net/ethernet/amazon/ena/ena_netdev.c
index 7eb6456ed0d5..80f5d08305c2 100644
--- a/drivers/net/ethernet/amazon/ena/ena_netdev.c
+++ b/drivers/net/ethernet/amazon/ena/ena_netdev.c
@@ -986,7 +986,7 @@ static struct sk_buff *ena_rx_skb(struct ena_ring *rx_ring,
page_offset = rx_info->page_offset;
buf_addr = page_address(rx_info->page) + page_offset;
- if (len <= rx_ring->rx_copybreak) {
+ if (len <= rx_ring->rx_copybreak && likely(descs == 1)) {
skb = ena_alloc_skb(rx_ring, NULL, len);
if (unlikely(!skb))
return NULL;@@ -1293,7 +1293,8 @@ static int ena_clean_rx_irq(struct ena_ring *rx_ring, struct napi_struct *napi,
skb_record_rx_queue(skb, rx_ring->qid);
- if (rx_ring->ena_bufs[0].len <= rx_ring->rx_copybreak)
+ if (rx_ring->ena_bufs[0].len <= rx_ring->rx_copybreak &&
+ ena_rx_ctx.descs == 1)
rx_copybreak_pkt++;
total_len += skb->len;
base-commit: 1d5e2c57e9ffc134e14b3f837c8d7c53b508bd22
--
2.43.0