[PATCH net-next 3/4] net: rmnet: conditionally expand skb headroom in ingress handler

Currently, when the rmnet driver processes ingress packets for an Ethernet
device (ARPHRD_ETHER), it unconditionally calls pskb_expand_head() to
ensure there is enough room to push the MAC header. This forces a costly
memory reallocation and payload copy for every single packet on the RX
hotpath, severely degrading throughput and increasing CPU overhead.

In many modern architectures, the underlying physical driver (e.g., USB)
can be configured to pre-allocate this extra ETH_HLEN headroom when
minting the initial SKB.

Optimize the ingress path by checking if skb_headroom(skb) < ETH_HLEN
before triggering the expansion. If the packet arrives with sufficient
headroom, the driver now skips the reallocation entirely and safely pushes
the header. The expensive pskb_expand_head() operation is now strictly a
fallback, allowing properly configured hardware to achieve zero-copy MAC
header insertion.

Testing this using a temporary print in the new condition showed that
the expansion is not triggered as enough space is already available.

Signed-off-by: Koen Vandeputte <redacted>
---
 drivers/net/ethernet/qualcomm/rmnet/rmnet_handlers.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/drivers/net/ethernet/qualcomm/rmnet/rmnet_handlers.c b/drivers/net/ethernet/qualcomm/rmnet/rmnet_handlers.c
index aa5523f4618e..95c3e3934fd3 100644
--- a/drivers/net/ethernet/qualcomm/rmnet/rmnet_handlers.c
+++ b/drivers/net/ethernet/qualcomm/rmnet/rmnet_handlers.c

@@ -114,9 +114,11 @@ rmnet_map_ingress_handler(struct sk_buff *skb,
 	u32 data_format;
 
 	if (skb->dev->type == ARPHRD_ETHER) {
-		if (pskb_expand_head(skb, ETH_HLEN, 0, GFP_ATOMIC)) {
-			kfree_skb(skb);
-			return;
+		if (skb_headroom(skb) < ETH_HLEN) {
+			if (pskb_expand_head(skb, ETH_HLEN, 0, GFP_ATOMIC)) {
+				kfree_skb(skb);
+				return;
+			}
 		}
 
 		skb_push(skb, ETH_HLEN);

-- 
2.43.0