[PATCH mptcp-next v10 3/4] mptcp: support MSG_ERRQUEUE on the parent socket

[David Carlier <devnexen@gmail.com>]

Splice pending err skbs from each subflow's error queue onto the parent
msk's error queue at error-report time, so poll() and recvmsg(MSG_ERRQUEUE)
on the parent socket observe TX timestamps and MSG_ZEROCOPY completion
notifications through the standard inet ABI.

The splice filters by SO_EE_ORIGIN: TIMESTAMPING / ZEROCOPY / LOCAL
events forward to the parent because they are tied to user-handed data,
not to a specific path; subflow-level ICMP errors are dropped because
the legacy RECVERR ABI cannot meaningfully convey their per-subflow peer
identity to single-path-aware userspace. Such events will be carried by
a future MPTCP_RECERR channel.

MSG_ZEROCOPY completions carry an unpin/free obligation for userspace, so
they are queued to the parent unconditionally and are never dropped,
mirroring tcp's __msg_zerocopy_callback() which likewise bypasses
sk_rcvbuf. Timestamping and local events instead go through
sock_queue_err_skb() and are dropped under rmem pressure (sk_rmem_alloc +
truesize >= sk_rcvbuf) on a full err queue, matching tcp's sk_rcvbuf-gated
tx-timestamp path and ip_icmp_error() / ipv6_icmp_error(). The
MSG_ERRQUEUE branch of mptcp_recvmsg() forwards to inet_recv_error()
directly, and poll() advertises EPOLLERR purely on the parent's sk_err /
sk_error_queue, matching tcp_poll().

Suggested-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: David Carlier <devnexen@gmail.com>
---
 net/mptcp/protocol.c | 68 ++++++++++++++++++++++++++++++++++++++------
 1 file changed, 59 insertions(+), 9 deletions(-)

diff --git a/net/mptcp/protocol.c b/net/mptcp/protocol.c
index 1d67728d4233..64c0e841b9b7 100644
--- a/net/mptcp/protocol.c
+++ b/net/mptcp/protocol.c
@@ -11,6 +11,7 @@
 #include <linux/netdevice.h>
 #include <linux/sched/signal.h>
 #include <linux/atomic.h>
+#include <linux/errqueue.h>
 #include <net/aligned_data.h>
 #include <net/rps.h>
 #include <net/sock.h>
@@ -829,21 +830,66 @@ static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
 	return moved;
 }
 
+static bool mptcp_errqueue_skb_forwardable(const struct sk_buff *skb)
+{
+	u8 origin = SKB_EXT_ERR(skb)->ee.ee_origin;
+
+	return origin == SO_EE_ORIGIN_TIMESTAMPING ||
+		origin == SO_EE_ORIGIN_ZEROCOPY ||
+		origin == SO_EE_ORIGIN_LOCAL;
+}
+
+static bool __mptcp_subflow_splice_errqueue(struct sock *sk, struct sock *ssk)
+{
+	struct sk_buff *skb;
+	bool moved = false;
+
+	while ((skb = skb_dequeue(&ssk->sk_error_queue))) {
+		if (!mptcp_errqueue_skb_forwardable(skb)) {
+			kfree_skb(skb);  /* path-specific (ICMP) — belongs in MPTCP_RECERR */
+			continue;
+		}
+		/* MSG_ZEROCOPY completions carry an unpin/free obligation for
+		 * userspace and must never be dropped. TCP's
+		 * __msg_zerocopy_callback() queues them to sk_error_queue
+		 * regardless of sk_rcvbuf; mirror that here rather than letting
+		 * sock_queue_err_skb() drop them under receive-buffer pressure.
+		 */
+		if (SKB_EXT_ERR(skb)->ee.ee_origin == SO_EE_ORIGIN_ZEROCOPY) {
+			skb_orphan(skb);
+			skb->sk = sk;
+			skb_queue_tail(&sk->sk_error_queue, skb);
+			moved = true;
+			continue;
+		}
+		if (sock_queue_err_skb(sk, skb)) {
+			kfree_skb(skb);
+			continue;
+		}
+		moved = true;
+	}
+
+	return moved;
+}
+
 static bool __mptcp_subflow_error_report(struct sock *sk, struct sock *ssk)
 {
+	bool propagated = false;
 	int ssk_state;
+	bool report;
 	int err;
 
+	report = __mptcp_subflow_splice_errqueue(sk, ssk);
+
 	/* only propagate errors on fallen-back sockets or
 	 * on MPC connect
 	 */
 	if (sk->sk_state != TCP_SYN_SENT && !__mptcp_check_fallback(mptcp_sk(sk)))
-		return false;
+		goto out;
 
 	err = sock_error(ssk);
 	if (!err)
-		return false;
-
+		goto out;
 	/* We need to propagate only transition to CLOSE state.
 	 * Orphaned socket will see such state change via
 	 * subflow_sched_work_if_closed() and that path will properly
@@ -853,11 +899,15 @@ static bool __mptcp_subflow_error_report(struct sock *sk, struct sock *ssk)
 	if (ssk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DEAD))
 		mptcp_set_state(sk, ssk_state);
 	WRITE_ONCE(sk->sk_err, -err);
+	report = propagated = true;
 
-	/* This barrier is coupled with smp_rmb() in mptcp_poll() */
-	smp_wmb();
-	sk_error_report(sk);
-	return true;
+out:
+	if (report) {
+		/* This barrier is coupled with smp_rmb() in mptcp_poll() */
+		smp_wmb();
+		sk_error_report(sk);
+	}
+	return propagated;
 }
 
 void __mptcp_error_report(struct sock *sk)
@@ -2313,7 +2363,6 @@ static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
 	int target;
 	long timeo;
 
-	/* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
 	if (unlikely(flags & MSG_ERRQUEUE))
 		return inet_recv_error(sk, msg, len);
 
@@ -4363,7 +4412,8 @@ static __poll_t mptcp_poll(struct file *file, struct socket *sock,
 
 	/* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
 	smp_rmb();
-	if (READ_ONCE(sk->sk_err))
+	if (READ_ONCE(sk->sk_err) ||
+	    !skb_queue_empty_lockless(&sk->sk_error_queue))
 		mask |= EPOLLERR;
 
 	return mask;
-- 
2.53.0