Merge branch 'mptcp-lowat-sockopt'

Matthieu Baerts says:

====================
mptcp: add TCP_NOTSENT_LOWAT sockopt support

Patch 3 does the magic of adding TCP_NOTSENT_LOWAT support, all the
other ones are minor cleanup seen along when working on the new
feature.

Note that this feature relies on the existing accounting for snd_nxt.
Such accounting is not 110% accurate as it tracks the most recent
sequence number queued to any subflow, and not the actual sequence
number sent on the wire. Paolo experimented a lot, trying to implement
the latter, and in the end it proved to be both "too complex" and "not
necessary".

The complexity raises from the need for additional lock and a lot of
refactoring to introduce such protections without adding significant
overhead. Additionally, snd_nxt is currently used and exposed with the
current semantic by the internal packet scheduling. Introducing a
different tracking will still require us to keep the old one.

More interestingly, a more accurate tracking could be not strictly
necessary: as the MPTCP socket enqueues data to the subflows only up
to the available send window, any enqueue data is sent on the wire
instantly, without any blocking operation short or a drop in the tx
path at the nft or TC layer.
====================
Signed-off-by: Matthieu Baerts (NGI0) <matttbe@kernel.org>

net/mptcp/protocol.c

@@ -1692,15 +1692,6 @@ static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool
 	}
 }
 
-static void mptcp_set_nospace(struct sock *sk)
-{
-	/* enable autotune */
-	set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
-
-	/* will be cleared on avail space */
-	set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
-}
-
 static int mptcp_disconnect(struct sock *sk, int flags);
 
 static int mptcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
@@ -1771,6 +1762,30 @@ static int do_copy_data_nocache(struct sock *sk, int copy,
 	return 0;
 }
 
+/* open-code sk_stream_memory_free() plus sent limit computation to
+ * avoid indirect calls in fast-path.
+ * Called under the msk socket lock, so we can avoid a bunch of ONCE
+ * annotations.
+ */
+static u32 mptcp_send_limit(const struct sock *sk)
+{
+	const struct mptcp_sock *msk = mptcp_sk(sk);
+	u32 limit, not_sent;
+
+	if (sk->sk_wmem_queued >= READ_ONCE(sk->sk_sndbuf))
+		return 0;
+
+	limit = mptcp_notsent_lowat(sk);
+	if (limit == UINT_MAX)
+		return UINT_MAX;
+
+	not_sent = msk->write_seq - msk->snd_nxt;
+	if (not_sent >= limit)
+		return 0;
+
+	return limit - not_sent;
+}
+
 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
 {
 	struct mptcp_sock *msk = mptcp_sk(sk);
@@ -1815,6 +1830,12 @@ static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
 		struct mptcp_data_frag *dfrag;
 		bool dfrag_collapsed;
 		size_t psize, offset;
+		u32 copy_limit;
+
+		/* ensure fitting the notsent_lowat() constraint */
+		copy_limit = mptcp_send_limit(sk);
+		if (!copy_limit)
+			goto wait_for_memory;
 
 		/* reuse tail pfrag, if possible, or carve a new one from the
 		 * page allocator
@@ -1822,9 +1843,6 @@ static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
 		dfrag = mptcp_pending_tail(sk);
 		dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
 		if (!dfrag_collapsed) {
-			if (!sk_stream_memory_free(sk))
-				goto wait_for_memory;
-
 			if (!mptcp_page_frag_refill(sk, pfrag))
 				goto wait_for_memory;
 
@@ -1839,6 +1857,7 @@ static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
 		offset = dfrag->offset + dfrag->data_len;
 		psize = pfrag->size - offset;
 		psize = min_t(size_t, psize, msg_data_left(msg));
+		psize = min_t(size_t, psize, copy_limit);
 		total_ts = psize + frag_truesize;
 
 		if (!sk_wmem_schedule(sk, total_ts))
@@ -1874,7 +1893,7 @@ static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
 		continue;
 
 wait_for_memory:
-		mptcp_set_nospace(sk);
+		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 		__mptcp_push_pending(sk, msg->msg_flags);
 		ret = sk_stream_wait_memory(sk, &timeo);
 		if (ret)
@@ -3769,6 +3788,7 @@ static struct proto mptcp_prot = {
 	.unhash		= mptcp_unhash,
 	.get_port	= mptcp_get_port,
 	.forward_alloc_get	= mptcp_forward_alloc_get,
+	.stream_memory_free	= mptcp_stream_memory_free,
 	.sockets_allocated	= &mptcp_sockets_allocated,
 
 	.memory_allocated	= &tcp_memory_allocated,
@@ -3942,12 +3962,12 @@ static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
 {
 	struct sock *sk = (struct sock *)msk;
 
-	if (sk_stream_is_writeable(sk))
+	if (__mptcp_stream_is_writeable(sk, 1))
 		return EPOLLOUT | EPOLLWRNORM;
 
-	mptcp_set_nospace(sk);
-	smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
-	if (sk_stream_is_writeable(sk))
+	set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
+	smp_mb__after_atomic(); /* NOSPACE is changed by mptcp_write_space() */
+	if (__mptcp_stream_is_writeable(sk, 1))
 		return EPOLLOUT | EPOLLWRNORM;
 
 	return 0;

net/mptcp/protocol.h

@@ -113,10 +113,9 @@
 #define MPTCP_RST_TRANSIENT	BIT(0)
 
 /* MPTCP socket atomic flags */
-#define MPTCP_NOSPACE		1
-#define MPTCP_WORK_RTX		2
-#define MPTCP_FALLBACK_DONE	4
-#define MPTCP_WORK_CLOSE_SUBFLOW 5
+#define MPTCP_WORK_RTX		1
+#define MPTCP_FALLBACK_DONE	2
+#define MPTCP_WORK_CLOSE_SUBFLOW 3
 
 /* MPTCP socket release cb flags */
 #define MPTCP_PUSH_PENDING	1
@@ -308,6 +307,7 @@ struct mptcp_sock {
 			in_accept_queue:1,
 			free_first:1,
 			rcvspace_init:1;
+	u32		notsent_lowat;
 	struct work_struct work;
 	struct sk_buff  *ooo_last_skb;
 	struct rb_root  out_of_order_queue;
@@ -808,14 +808,36 @@ static inline bool mptcp_data_fin_enabled(const struct mptcp_sock *msk)
 	       READ_ONCE(msk->write_seq) == READ_ONCE(msk->snd_nxt);
 }
 
+static inline u32 mptcp_notsent_lowat(const struct sock *sk)
+{
+	struct net *net = sock_net(sk);
+	u32 val;
+
+	val = READ_ONCE(mptcp_sk(sk)->notsent_lowat);
+	return val ?: READ_ONCE(net->ipv4.sysctl_tcp_notsent_lowat);
+}
+
+static inline bool mptcp_stream_memory_free(const struct sock *sk, int wake)
+{
+	const struct mptcp_sock *msk = mptcp_sk(sk);
+	u32 notsent_bytes;
+
+	notsent_bytes = READ_ONCE(msk->write_seq) - READ_ONCE(msk->snd_nxt);
+	return (notsent_bytes << wake) < mptcp_notsent_lowat(sk);
+}
+
+static inline bool __mptcp_stream_is_writeable(const struct sock *sk, int wake)
+{
+	return mptcp_stream_memory_free(sk, wake) &&
+	       __sk_stream_is_writeable(sk, wake);
+}
+
 static inline void mptcp_write_space(struct sock *sk)
 {
-	if (sk_stream_is_writeable(sk)) {
-		/* pairs with memory barrier in mptcp_poll */
-		smp_mb();
-		if (test_and_clear_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags))
-			sk_stream_write_space(sk);
-	}
+	/* pairs with memory barrier in mptcp_poll */
+	smp_mb();
+	if (mptcp_stream_memory_free(sk, 1))
+		sk_stream_write_space(sk);
 }
 
 static inline void __mptcp_sync_sndbuf(struct sock *sk)

net/mptcp/sockopt.c

@@ -624,20 +624,11 @@ static int mptcp_setsockopt_sol_tcp_congestion(struct mptcp_sock *msk, sockptr_t
 	return ret;
 }
 
-static int mptcp_setsockopt_sol_tcp_cork(struct mptcp_sock *msk, sockptr_t optval,
-					 unsigned int optlen)
+static int __mptcp_setsockopt_sol_tcp_cork(struct mptcp_sock *msk, int val)
 {
 	struct mptcp_subflow_context *subflow;
 	struct sock *sk = (struct sock *)msk;
-	int val;
 
-	if (optlen < sizeof(int))
-		return -EINVAL;
-
-	if (copy_from_sockptr(&val, optval, sizeof(val)))
-		return -EFAULT;
-
-	lock_sock(sk);
 	sockopt_seq_inc(msk);
 	msk->cork = !!val;
 	mptcp_for_each_subflow(msk, subflow) {
@@ -649,25 +640,15 @@ static int mptcp_setsockopt_sol_tcp_cork(struct mptcp_sock *msk, sockptr_t optva
 	}
 	if (!val)
 		mptcp_check_and_set_pending(sk);
-	release_sock(sk);
 
 	return 0;
 }
 
-static int mptcp_setsockopt_sol_tcp_nodelay(struct mptcp_sock *msk, sockptr_t optval,
-					    unsigned int optlen)
+static int __mptcp_setsockopt_sol_tcp_nodelay(struct mptcp_sock *msk, int val)
 {
 	struct mptcp_subflow_context *subflow;
 	struct sock *sk = (struct sock *)msk;
-	int val;
-
-	if (optlen < sizeof(int))
-		return -EINVAL;
-
-	if (copy_from_sockptr(&val, optval, sizeof(val)))
-		return -EFAULT;
 
-	lock_sock(sk);
 	sockopt_seq_inc(msk);
 	msk->nodelay = !!val;
 	mptcp_for_each_subflow(msk, subflow) {
@@ -679,8 +660,6 @@ static int mptcp_setsockopt_sol_tcp_nodelay(struct mptcp_sock *msk, sockptr_t op
 	}
 	if (val)
 		mptcp_check_and_set_pending(sk);
-	release_sock(sk);
-
 	return 0;
 }
 
@@ -803,25 +782,10 @@ static int mptcp_setsockopt_sol_tcp(struct mptcp_sock *msk, int optname,
 	int ret, val;
 
 	switch (optname) {
-	case TCP_INQ:
-		ret = mptcp_get_int_option(msk, optval, optlen, &val);
-		if (ret)
-			return ret;
-		if (val < 0 || val > 1)
-			return -EINVAL;
-
-		lock_sock(sk);
-		msk->recvmsg_inq = !!val;
-		release_sock(sk);
-		return 0;
 	case TCP_ULP:
 		return -EOPNOTSUPP;
 	case TCP_CONGESTION:
 		return mptcp_setsockopt_sol_tcp_congestion(msk, optval, optlen);
-	case TCP_CORK:
-		return mptcp_setsockopt_sol_tcp_cork(msk, optval, optlen);
-	case TCP_NODELAY:
-		return mptcp_setsockopt_sol_tcp_nodelay(msk, optval, optlen);
 	case TCP_DEFER_ACCEPT:
 		/* See tcp.c: TCP_DEFER_ACCEPT does not fail */
 		mptcp_setsockopt_first_sf_only(msk, SOL_TCP, optname, optval, optlen);
@@ -834,7 +798,34 @@ static int mptcp_setsockopt_sol_tcp(struct mptcp_sock *msk, int optname,
 						      optval, optlen);
 	}
 
-	return -EOPNOTSUPP;
+	ret = mptcp_get_int_option(msk, optval, optlen, &val);
+	if (ret)
+		return ret;
+
+	lock_sock(sk);
+	switch (optname) {
+	case TCP_INQ:
+		if (val < 0 || val > 1)
+			ret = -EINVAL;
+		else
+			msk->recvmsg_inq = !!val;
+		break;
+	case TCP_NOTSENT_LOWAT:
+		WRITE_ONCE(msk->notsent_lowat, val);
+		mptcp_write_space(sk);
+		break;
+	case TCP_CORK:
+		ret = __mptcp_setsockopt_sol_tcp_cork(msk, val);
+		break;
+	case TCP_NODELAY:
+		ret = __mptcp_setsockopt_sol_tcp_nodelay(msk, val);
+		break;
+	default:
+		ret = -ENOPROTOOPT;
+	}
+
+	release_sock(sk);
+	return ret;
 }
 
 int mptcp_setsockopt(struct sock *sk, int level, int optname,
@@ -1349,6 +1340,8 @@ static int mptcp_getsockopt_sol_tcp(struct mptcp_sock *msk, int optname,
 		return mptcp_put_int_option(msk, optval, optlen, msk->cork);
 	case TCP_NODELAY:
 		return mptcp_put_int_option(msk, optval, optlen, msk->nodelay);
+	case TCP_NOTSENT_LOWAT:
+		return mptcp_put_int_option(msk, optval, optlen, msk->notsent_lowat);
 	}
 	return -EOPNOTSUPP;
 }