kcm: fix zero-frag skb in frag_list on partial sendmsg error

Syzkaller reported a warning in kcm_write_msgs() when processing a
message with a zero-fragment skb in the frag_list.

When kcm_sendmsg() fills MAX_SKB_FRAGS fragments in the current skb,
it allocates a new skb (tskb) and links it into the frag_list before
copying data. If the copy subsequently fails (e.g. -EFAULT from
user memory), tskb remains in the frag_list with zero fragments:

  head skb (msg being assembled, NOT yet in sk_write_queue)
  +-----------+
  | frags[17] |  (MAX_SKB_FRAGS, all filled with data)
  | frag_list-+--> tskb
  +-----------+    +----------+
                   | frags[0] |  (empty! copy failed before filling)
                   +----------+

For SOCK_SEQPACKET with partial data already copied, the error path
saves this message via partial_message for later completion. For
SOCK_SEQPACKET, sock_write_iter() automatically sets MSG_EOR, so a
subsequent zero-length write(fd, NULL, 0) completes the message and
queues it to sk_write_queue. kcm_write_msgs() then walks the
frag_list and hits:

  WARN_ON(!skb_shinfo(skb)->nr_frags)

TCP has a similar pattern where skbs are enqueued before data copy
and cleaned up on failure via tcp_remove_empty_skb(). KCM was
missing the equivalent cleanup.

Fix this by tracking the predecessor skb (frag_prev) when allocating
a new frag_list entry. On error, if the tail skb has zero frags,
use frag_prev to unlink and free it in O(1) without walking the
singly-linked frag_list. frag_prev is safe to dereference because
the entire message chain is only held locally (or in kcm->seq_skb)
and is not added to sk_write_queue until MSG_EOR, so the send path
cannot free it underneath us.

Also change the WARN_ON to WARN_ON_ONCE to avoid flooding the log
if the condition is somehow hit repeatedly.

There are currently no KCM selftests in the kernel tree; a simple
reproducer is available at [1].

[1] https://gist.github.com/mrpre/a94d431c757e8d6f168f4dd1a3749daa

Reported-by: syzbot+52624bdfbf2746d37d70@syzkaller.appspotmail.com
Closes: https://lore.kernel.org/all/000000000000269a1405a12fdc77@google.com/T/
Fixes: ab7ac4eb9832 ("kcm: Kernel Connection Multiplexor module")
Signed-off-by: Jiayuan Chen <jiayuan.chen@shopee.com>
Link: https://patch.msgid.link/20260219014256.370092-1-jiayuan.chen@linux.dev
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

1 files changed, 19 insertions, 2 deletions

diff --git a/net/kcm/kcmsock.c b/net/kcm/kcmsock.c
index 5dd7e0509a48..3912e75079f5 100644
--- a/net/kcm/kcmsock.c
+++ b/net/kcm/kcmsock.c
@@ -628,7 +628,7 @@ retry:
 			skb = txm->frag_skb;
 		}
 
-		if (WARN_ON(!skb_shinfo(skb)->nr_frags) ||
+		if (WARN_ON_ONCE(!skb_shinfo(skb)->nr_frags) ||
 		    WARN_ON_ONCE(!skb_frag_page(&skb_shinfo(skb)->frags[0]))) {
 			ret = -EINVAL;
 			goto out;
@@ -749,7 +749,7 @@ static int kcm_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
 {
 	struct sock *sk = sock->sk;
 	struct kcm_sock *kcm = kcm_sk(sk);
-	struct sk_buff *skb = NULL, *head = NULL;
+	struct sk_buff *skb = NULL, *head = NULL, *frag_prev = NULL;
 	size_t copy, copied = 0;
 	long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
 	int eor = (sock->type == SOCK_DGRAM) ?
@@ -824,6 +824,7 @@ start:
 				else
 					skb->next = tskb;
 
+				frag_prev = skb;
 				skb = tskb;
 				skb->ip_summed = CHECKSUM_UNNECESSARY;
 				continue;
@@ -933,6 +934,22 @@ partial_message:
 out_error:
 	kcm_push(kcm);
 
+	/* When MAX_SKB_FRAGS was reached, a new skb was allocated and
+	 * linked into the frag_list before data copy. If the copy
+	 * subsequently failed, this skb has zero frags. Remove it from
+	 * the frag_list to prevent kcm_write_msgs from later hitting
+	 * WARN_ON(!skb_shinfo(skb)->nr_frags).
+	 */
+	if (frag_prev && !skb_shinfo(skb)->nr_frags) {
+		if (head == frag_prev)
+			skb_shinfo(head)->frag_list = NULL;
+		else
+			frag_prev->next = NULL;
+		kfree_skb(skb);
+		/* Update skb as it may be saved in partial_message via goto */
+		skb = frag_prev;
+	}
+
 	if (sock->type == SOCK_SEQPACKET) {
 		/* Wrote some bytes before encountering an
 		 * error, return partial success.