net: mctp: avoid copy in fragmentation loop for near-MTU messages

Currently, we incorrectly send messages that are within 4 bytes (a
struct mctp_hdr) smaller than the MTU through mctp_do_fragment_route().
This has no effect on the actual fragmentation, as we will still send as
one packet, but unnecessarily copies the original skb into a new
single-fragment skb.

Instead of having the MTU comparisons in both mctp_local_output() and
mctp_do_fragment_route(), feed all local messages through the latter,
and add the single-packet optimisation there.

This means we can coalesce the routing path of mctp_local_output, so our
out_release path is now solely for errors, so rename the label
accordingly.

Include a check in the route tests for the single-packet case too.

Reported-by: yuanzhaoming <yuanzm2@lenovo.com>
Closes: https://github.com/openbmc/linux/commit/269936db5eb3962fe290b1dc4dbf1859cd5a04dd#r175836230
Signed-off-by: Jeremy Kerr <jk@codeconstruct.com.au>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20260324-dev-mtu-copy-v1-1-7af6bd7027d3@codeconstruct.com.au
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

2 files changed, 15 insertions, 15 deletions

diff --git a/net/mctp/route.c b/net/mctp/route.c
index 59ad60b88563..021e04f1ea7c 100644
--- a/net/mctp/route.c
+++ b/net/mctp/route.c
@@ -1037,6 +1037,13 @@ static int mctp_do_fragment_route(struct mctp_dst *dst, struct sk_buff *skb,
 		return -EMSGSIZE;
 	}
 
+	/* within MTU? avoid the copy, send original skb */
+	if (skb->len <= mtu) {
+		hdr->flags_seq_tag = MCTP_HDR_FLAG_SOM |
+			MCTP_HDR_FLAG_EOM | tag;
+		return dst->output(dst, skb);
+	}
+
 	/* keep same headroom as the original skb */
 	headroom = skb_headroom(skb);
 
@@ -1111,7 +1118,6 @@ int mctp_local_output(struct sock *sk, struct mctp_dst *dst,
 	struct mctp_hdr *hdr;
 	unsigned long flags;
 	unsigned int netid;
-	unsigned int mtu;
 	mctp_eid_t saddr;
 	int rc;
 	u8 tag;
@@ -1133,7 +1139,7 @@ int mctp_local_output(struct sock *sk, struct mctp_dst *dst,
 	netid = READ_ONCE(dst->dev->net);
 
 	if (rc)
-		goto out_release;
+		goto err_free;
 
 	if (req_tag & MCTP_TAG_OWNER) {
 		if (req_tag & MCTP_TAG_PREALLOC)
@@ -1145,7 +1151,7 @@ int mctp_local_output(struct sock *sk, struct mctp_dst *dst,
 
 		if (IS_ERR(key)) {
 			rc = PTR_ERR(key);
-			goto out_release;
+			goto err_free;
 		}
 		mctp_skb_set_flow(skb, key);
 		/* done with the key in this scope */
@@ -1170,20 +1176,10 @@ int mctp_local_output(struct sock *sk, struct mctp_dst *dst,
 	hdr->dest = daddr;
 	hdr->src = saddr;
 
-	mtu = dst->mtu;
-
-	if (skb->len + sizeof(struct mctp_hdr) <= mtu) {
-		hdr->flags_seq_tag = MCTP_HDR_FLAG_SOM |
-			MCTP_HDR_FLAG_EOM | tag;
-		rc = dst->output(dst, skb);
-	} else {
-		rc = mctp_do_fragment_route(dst, skb, mtu, tag);
-	}
-
 	/* route output functions consume the skb, even on error */
-	skb = NULL;
+	return mctp_do_fragment_route(dst, skb, dst->mtu, tag);
 
-out_release:
+err_free:
 	kfree_skb(skb);
 	return rc;
 }
diff --git a/net/mctp/test/route-test.c b/net/mctp/test/route-test.c
index 75ea96c10e49..61c989c43ec0 100644
--- a/net/mctp/test/route-test.c
+++ b/net/mctp/test/route-test.c
@@ -63,6 +63,10 @@ static void mctp_test_fragment(struct kunit *test)
 		if (!skb2)
 			break;
 
+		/* avoid copying single-skb messages */
+		if (first && last)
+			KUNIT_EXPECT_PTR_EQ(test, skb, skb2);
+
 		hdr2 = mctp_hdr(skb2);
 
 		tag_mask = MCTP_HDR_TAG_MASK | MCTP_HDR_FLAG_TO;