[PATCH net-next 0/2] Add IP_LOCAL_PORT_RANGE socket option

[PATCH net-next 0/2] Add IP_LOCAL_PORT_RANGE socket option

[Jakub Sitnicki]

This patch set is a follow up to the "How to share IPv4 addresses by
partitioning the port space" talk given at LPC 2022 [1].

Please see patch #1 for the motivation & the use case description.
Patch #2 adds tests exercising the new option in various scenarios.

If the changes were to be accepted, here is the proposed update to the ip(7)
man-page:

  IP_LOCAL_PORT_RANGE (since Linux X.Y)
         Set  or get the per-socket default local port range. This option
         can be used to clamp down the global local port  range,  defined
         by  the ip_local_port_range /proc interface described below, for
         a socket. The option takes a uint32_t value  with  the  high  16
         bits  set  to  the upper range bound, and the low 16 bits set to
         the lower range bound. Range bounds are inclusive. If the speci‐
         fied  high  or  low  bound  is  outside of the global local port
         range, or is set to zero, the set bound has no effect.

Changelog:
---------
RFC -> v1
RFC: https://lore.kernel.org/netdev/20220912225308.93659-1-jakub@cloudflare.com/

 * Allow either the high bound or the low bound, or both, to be zero
 * Add getsockopt support
 * Add selftests

Links:
------
[1]: https://lpc.events/event/16/contributions/1349/

To: netdev@vger.kernel.org
To: "David S. Miller" <davem@davemloft.net>
To: Eric Dumazet <edumazet@google.com>
To: Jakub Kicinski <kuba@kernel.org>
To: Paolo Abeni <pabeni@redhat.com>
Cc: kernel-team@cloudflare.com
Signed-off-by: Jakub Sitnicki <jakub@cloudflare.com>

---
Jakub Sitnicki (2):
      inet: Add IP_LOCAL_PORT_RANGE socket option
      selftests/net: Cover the IP_LOCAL_PORT_RANGE socket option

 include/net/inet_sock.h                            |   4 +
 include/net/ip.h                                   |   3 +-
 include/uapi/linux/in.h                            |   1 +
 net/ipv4/inet_connection_sock.c                    |  22 +-
 net/ipv4/inet_hashtables.c                         |   2 +-
 net/ipv4/ip_sockglue.c                             |  18 +
 net/ipv4/udp.c                                     |   2 +-
 tools/testing/selftests/net/Makefile               |   2 +
 tools/testing/selftests/net/ip_local_port_range.c  | 439 +++++++++++++++++++++
 tools/testing/selftests/net/ip_local_port_range.sh |   5 +
 10 files changed, 493 insertions(+), 5 deletions(-)
---
base-commit: 3d759e9e24c38758abc19a4f5e1872a6460d5745
change-id: 20221221-sockopt-port-range-e142de700f4d

Best regards,
-- 
Jakub Sitnicki <jakub@cloudflare.com>

[PATCH net-next 1/2] inet: Add IP_LOCAL_PORT_RANGE socket option

[Jakub Sitnicki]

Users who want to share a single public IP address for outgoing connections
between several hosts traditionally reach for SNAT. However, SNAT requires
state keeping on the node(s) performing the NAT.

A stateless alternative exists, where a single IP address used for egress
can be shared between several hosts by partitioning the available ephemeral
port range. In such a setup:

1. Each host gets assigned a disjoint range of ephemeral ports.
2. Applications open connections from the host-assigned port range.
3. Return traffic gets routed to the host based on both, the destination IP
   and the destination port.

An application which wants to open an outgoing connection (connect) from a
given port range today can choose between two solutions:

1. Manually pick the source port by bind()'ing to it before connect()'ing
   the socket.

   This approach has a couple of downsides:

   a) Search for a free port has to be implemented in the user-space. If
      the chosen 4-tuple happens to be busy, the application needs to retry
      from a different local port number.

      Detecting if 4-tuple is busy can be either easy (TCP) or hard
      (UDP). In TCP case, the application simply has to check if connect()
      returned an error (EADDRNOTAVAIL). That is assuming that the local
      port sharing was enabled (REUSEADDR) by all the sockets.

        # Assume desired local port range is 60_000-60_511
        s = socket(AF_INET, SOCK_STREAM)
        s.setsockopt(SOL_SOCKET, SO_REUSEADDR, 1)
        s.bind(("192.0.2.1", 60_000))
        s.connect(("1.1.1.1", 53))
        # Fails only if 192.0.2.1:60000 -> 1.1.1.1:53 is busy
        # Application must retry with another local port

      In case of UDP, the network stack allows binding more than one socket
      to the same 4-tuple, when local port sharing is enabled
      (REUSEADDR). Hence detecting the conflict is much harder and involves
      querying sock_diag and toggling the REUSEADDR flag [1].

   b) For TCP, bind()-ing to a port within the ephemeral port range means
      that no connecting sockets, that is those which leave it to the
      network stack to find a free local port at connect() time, can use
      the this port.

      IOW, the bind hash bucket tb->fastreuse will be 0 or 1, and the port
      will be skipped during the free port search at connect() time.

2. Isolate the app in a dedicated netns and use the use the per-netns
   ip_local_port_range sysctl to adjust the ephemeral port range bounds.

   The per-netns setting affects all sockets, so this approach can be used
   only if:

   - there is just one egress IP address, or
   - the desired egress port range is the same for all egress IP addresses
     used by the application.

   For TCP, this approach avoids the downsides of (1). Free port search and
   4-tuple conflict detection is done by the network stack:

     system("sysctl -w net.ipv4.ip_local_port_range='60000 60511'")

     s = socket(AF_INET, SOCK_STREAM)
     s.setsockopt(SOL_IP, IP_BIND_ADDRESS_NO_PORT, 1)
     s.bind(("192.0.2.1", 0))
     s.connect(("1.1.1.1", 53))
     # Fails if all 4-tuples 192.0.2.1:60000-60511 -> 1.1.1.1:53 are busy

  For UDP this approach has limited applicability. Setting the
  IP_BIND_ADDRESS_NO_PORT socket option does not result in local source
  port being shared with other connected UDP sockets.

  Hence relying on the network stack to find a free source port, limits the
  number of outgoing UDP flows from a single IP address down to the number
  of available ephemeral ports.

To put it another way, partitioning the ephemeral port range between hosts
using the existing Linux networking API is cumbersome.

To address this use case, add a new socket option at the SOL_IP level,
named IP_LOCAL_PORT_RANGE. The new option can be used to clamp down the
ephemeral port range for each socket individually.

The option can be used only to narrow down the per-netns local port
range. If the per-socket range lies outside of the per-netns range, the
latter takes precedence.

UAPI-wise, the low and high range bounds are passed to the kernel as a pair
of u16 values packed into a u32. This avoids pointer passing.

  PORT_LO = 40_000
  PORT_HI = 40_511

  s = socket(AF_INET, SOCK_STREAM)
  v = struct.pack("I", PORT_LO | (PORT_HI << 16))
  s.setsockopt(SOL_IP, IP_LOCAL_PORT_RANGE, v)
  s.bind(("127.0.0.1", 0))
  s.getsockname()
  # Local address between ("127.0.0.1", 40_000) and ("127.0.0.1", 40_511),
  # if there is a free port. EADDRINUSE otherwise.

[1] https://github.com/cloudflare/cloudflare-blog/blob/232b432c1d57/2022-02-connectx/connectx.py#L116

Reviewed-by: Marek Majkowski <marek@cloudflare.com>
Signed-off-by: Jakub Sitnicki <jakub@cloudflare.com>
---
 include/net/inet_sock.h         |  4 ++++
 include/net/ip.h                |  3 ++-
 include/uapi/linux/in.h         |  1 +
 net/ipv4/inet_connection_sock.c | 22 ++++++++++++++++++++--
 net/ipv4/inet_hashtables.c      |  2 +-
 net/ipv4/ip_sockglue.c          | 18 ++++++++++++++++++
 net/ipv4/udp.c                  |  2 +-
 7 files changed, 47 insertions(+), 5 deletions(-)

diff --git a/include/net/inet_sock.h b/include/net/inet_sock.h
index bf5654ce711e..51857117ac09 100644
--- a/include/net/inet_sock.h
+++ b/include/net/inet_sock.h
@@ -249,6 +249,10 @@ struct inet_sock {
 	__be32			mc_addr;
 	struct ip_mc_socklist __rcu	*mc_list;
 	struct inet_cork_full	cork;
+	struct {
+		__u16 lo;
+		__u16 hi;
+	}			local_port_range;
 };
 
 #define IPCORK_OPT	1	/* ip-options has been held in ipcork.opt */
diff --git a/include/net/ip.h b/include/net/ip.h
index 144bdfbb25af..c3fffaa92d6e 100644
--- a/include/net/ip.h
+++ b/include/net/ip.h
@@ -340,7 +340,8 @@ static inline u64 snmp_fold_field64(void __percpu *mib, int offt, size_t syncp_o
 	} \
 }
 
-void inet_get_local_port_range(struct net *net, int *low, int *high);
+void inet_get_local_port_range(const struct net *net, int *low, int *high);
+void inet_sk_get_local_port_range(const struct sock *sk, int *low, int *high);
 
 #ifdef CONFIG_SYSCTL
 static inline bool inet_is_local_reserved_port(struct net *net, unsigned short port)
diff --git a/include/uapi/linux/in.h b/include/uapi/linux/in.h
index 07a4cb149305..4b7f2df66b99 100644
--- a/include/uapi/linux/in.h
+++ b/include/uapi/linux/in.h
@@ -162,6 +162,7 @@ struct in_addr {
 #define MCAST_MSFILTER			48
 #define IP_MULTICAST_ALL		49
 #define IP_UNICAST_IF			50
+#define IP_LOCAL_PORT_RANGE		51
 
 #define MCAST_EXCLUDE	0
 #define MCAST_INCLUDE	1
diff --git a/net/ipv4/inet_connection_sock.c b/net/ipv4/inet_connection_sock.c
index d1f837579398..ec40c9ee753c 100644
--- a/net/ipv4/inet_connection_sock.c
+++ b/net/ipv4/inet_connection_sock.c
@@ -117,7 +117,7 @@ bool inet_rcv_saddr_any(const struct sock *sk)
 	return !sk->sk_rcv_saddr;
 }
 
-void inet_get_local_port_range(struct net *net, int *low, int *high)
+void inet_get_local_port_range(const struct net *net, int *low, int *high)
 {
 	unsigned int seq;
 
@@ -130,6 +130,24 @@ void inet_get_local_port_range(struct net *net, int *low, int *high)
 }
 EXPORT_SYMBOL(inet_get_local_port_range);
 
+void inet_sk_get_local_port_range(const struct sock *sk, int *low, int *high)
+{
+	const struct inet_sock *inet = inet_sk(sk);
+	const struct net *net = sock_net(sk);
+	int lo, hi;
+
+	inet_get_local_port_range(net, &lo, &hi);
+
+	if (unlikely(inet->local_port_range.lo))
+		lo = clamp_val(inet->local_port_range.lo, lo, hi);
+	if (unlikely(inet->local_port_range.hi))
+		hi = clamp_val(inet->local_port_range.hi, lo, hi);
+
+	*low = lo;
+	*high = hi;
+}
+EXPORT_SYMBOL(inet_sk_get_local_port_range);
+
 static bool inet_use_bhash2_on_bind(const struct sock *sk)
 {
 #if IS_ENABLED(CONFIG_IPV6)
@@ -316,7 +334,7 @@ inet_csk_find_open_port(const struct sock *sk, struct inet_bind_bucket **tb_ret,
 ports_exhausted:
 	attempt_half = (sk->sk_reuse == SK_CAN_REUSE) ? 1 : 0;
 other_half_scan:
-	inet_get_local_port_range(net, &low, &high);
+	inet_sk_get_local_port_range(sk, &low, &high);
 	high++; /* [32768, 60999] -> [32768, 61000[ */
 	if (high - low < 4)
 		attempt_half = 0;
diff --git a/net/ipv4/inet_hashtables.c b/net/ipv4/inet_hashtables.c
index 24a38b56fab9..22ee5f23ee5b 100644
--- a/net/ipv4/inet_hashtables.c
+++ b/net/ipv4/inet_hashtables.c
@@ -1013,7 +1013,7 @@ int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 
 	l3mdev = inet_sk_bound_l3mdev(sk);
 
-	inet_get_local_port_range(net, &low, &high);
+	inet_sk_get_local_port_range(sk, &low, &high);
 	high++; /* [32768, 60999] -> [32768, 61000[ */
 	remaining = high - low;
 	if (likely(remaining > 1))
diff --git a/net/ipv4/ip_sockglue.c b/net/ipv4/ip_sockglue.c
index 9f92ae35bb01..b511ff0adc0a 100644
--- a/net/ipv4/ip_sockglue.c
+++ b/net/ipv4/ip_sockglue.c
@@ -923,6 +923,7 @@ int do_ip_setsockopt(struct sock *sk, int level, int optname,
 	case IP_CHECKSUM:
 	case IP_RECVFRAGSIZE:
 	case IP_RECVERR_RFC4884:
+	case IP_LOCAL_PORT_RANGE:
 		if (optlen >= sizeof(int)) {
 			if (copy_from_sockptr(&val, optval, sizeof(val)))
 				return -EFAULT;
@@ -1365,6 +1366,20 @@ int do_ip_setsockopt(struct sock *sk, int level, int optname,
 		WRITE_ONCE(inet->min_ttl, val);
 		break;
 
+	case IP_LOCAL_PORT_RANGE:
+	{
+		const __u16 lo = val;
+		const __u16 hi = val >> 16;
+
+		if (optlen != sizeof(__u32))
+			goto e_inval;
+		if (lo != 0 && hi != 0 && lo > hi)
+			goto e_inval;
+
+		inet->local_port_range.lo = lo;
+		inet->local_port_range.hi = hi;
+		break;
+	}
 	default:
 		err = -ENOPROTOOPT;
 		break;
@@ -1743,6 +1758,9 @@ int do_ip_getsockopt(struct sock *sk, int level, int optname,
 	case IP_MINTTL:
 		val = inet->min_ttl;
 		break;
+	case IP_LOCAL_PORT_RANGE:
+		val = inet->local_port_range.hi << 16 | inet->local_port_range.lo;
+		break;
 	default:
 		sockopt_release_sock(sk);
 		return -ENOPROTOOPT;
diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c
index 9592fe3e444a..c605d171eb2d 100644
--- a/net/ipv4/udp.c
+++ b/net/ipv4/udp.c
@@ -248,7 +248,7 @@ int udp_lib_get_port(struct sock *sk, unsigned short snum,
 		int low, high, remaining;
 		unsigned int rand;
 
-		inet_get_local_port_range(net, &low, &high);
+		inet_sk_get_local_port_range(sk, &low, &high);
 		remaining = (high - low) + 1;
 
 		rand = get_random_u32();

-- 
2.38.1

[PATCH net-next 2/2] selftests/net: Cover the IP_LOCAL_PORT_RANGE socket option

[Jakub Sitnicki]

Exercise IP_LOCAL_PORT_RANGE socket option in various scenarios:

1. pass invalid values to setsockopt
2. pass a range outside of the per-netns port range
3. configure a single-port range
4. exhaust a configured multi-port range
5. check interaction with late-bind (IP_BIND_ADDRESS_NO_PORT)
6. set then get the per-socket port range

Signed-off-by: Jakub Sitnicki <jakub@cloudflare.com>
---
 tools/testing/selftests/net/Makefile               |   2 +
 tools/testing/selftests/net/ip_local_port_range.c  | 439 +++++++++++++++++++++
 tools/testing/selftests/net/ip_local_port_range.sh |   5 +
 3 files changed, 446 insertions(+)

diff --git a/tools/testing/selftests/net/Makefile b/tools/testing/selftests/net/Makefile
index 3007e98a6d64..51b0b49ecd3b 100644
--- a/tools/testing/selftests/net/Makefile
+++ b/tools/testing/selftests/net/Makefile
@@ -45,6 +45,7 @@ TEST_PROGS += arp_ndisc_untracked_subnets.sh
 TEST_PROGS += stress_reuseport_listen.sh
 TEST_PROGS += l2_tos_ttl_inherit.sh
 TEST_PROGS += bind_bhash.sh
+TEST_PROGS += ip_local_port_range.sh
 TEST_PROGS_EXTENDED := in_netns.sh setup_loopback.sh setup_veth.sh
 TEST_PROGS_EXTENDED += toeplitz_client.sh toeplitz.sh
 TEST_GEN_FILES =  socket nettest
@@ -75,6 +76,7 @@ TEST_GEN_PROGS += so_incoming_cpu
 TEST_PROGS += sctp_vrf.sh
 TEST_GEN_FILES += sctp_hello
 TEST_GEN_FILES += csum
+TEST_GEN_FILES += ip_local_port_range
 
 TEST_FILES := settings
 
diff --git a/tools/testing/selftests/net/ip_local_port_range.c b/tools/testing/selftests/net/ip_local_port_range.c
new file mode 100644
index 000000000000..272b0cb9fc57
--- /dev/null
+++ b/tools/testing/selftests/net/ip_local_port_range.c
@@ -0,0 +1,439 @@
+// SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
+// Copyright (c) 2022 Cloudflare
+
+/* Test IP_LOCAL_PORT_RANGE socket option: IPv4 + IPv6, TCP + UDP.
+ *
+ * Tests assume that net.ipv4.ip_local_port_range is [40000, 49999].
+ * Don't run these directly but with ip_local_port_range.sh script.
+ */
+
+#include <fcntl.h>
+#include <netinet/ip.h>
+
+#include "../kselftest_harness.h"
+
+#ifndef IP_LOCAL_PORT_RANGE
+#define IP_LOCAL_PORT_RANGE 51
+#endif
+
+static __u32 pack_port_range(__u16 lo, __u16 hi)
+{
+	return (hi << 16) | (lo << 0);
+}
+
+static void unpack_port_range(__u32 range, __u16 *lo, __u16 *hi)
+{
+	*lo = range & 0xffff;
+	*hi = range >> 16;
+}
+
+static int get_so_domain(int fd)
+{
+	int domain, err;
+	socklen_t len;
+
+	len = sizeof(domain);
+	err = getsockopt(fd, SOL_SOCKET, SO_DOMAIN, &domain, &len);
+	if (err)
+		return -1;
+
+	return domain;
+}
+
+static int bind_to_loopback_any_port(int fd)
+{
+	union {
+		struct sockaddr sa;
+		struct sockaddr_in v4;
+		struct sockaddr_in6 v6;
+	} addr;
+	socklen_t addr_len;
+
+	memset(&addr, 0, sizeof(addr));
+	switch (get_so_domain(fd)) {
+	case AF_INET:
+		addr.v4.sin_family = AF_INET;
+		addr.v4.sin_port = htons(0);
+		addr.v4.sin_addr.s_addr = htonl(INADDR_LOOPBACK);
+		addr_len = sizeof(addr.v4);
+		break;
+	case AF_INET6:
+		addr.v6.sin6_family = AF_INET6;
+		addr.v6.sin6_port = htons(0);
+		addr.v6.sin6_addr = in6addr_loopback;
+		addr_len = sizeof(addr.v6);
+		break;
+	default:
+		return -1;
+	}
+
+	return bind(fd, &addr.sa, addr_len);
+}
+
+static int get_sock_port(int fd)
+{
+	union {
+		struct sockaddr sa;
+		struct sockaddr_in v4;
+		struct sockaddr_in6 v6;
+	} addr;
+	socklen_t addr_len;
+	int err;
+
+	addr_len = sizeof(addr);
+	memset(&addr, 0, sizeof(addr));
+	err = getsockname(fd, &addr.sa, &addr_len);
+	if (err)
+		return -1;
+
+	switch (addr.sa.sa_family) {
+	case AF_INET:
+		return ntohs(addr.v4.sin_port);
+	case AF_INET6:
+		return ntohs(addr.v6.sin6_port);
+	default:
+		errno = EAFNOSUPPORT;
+		return -1;
+	}
+}
+
+static int get_ip_local_port_range(int fd, __u32 *range)
+{
+	socklen_t len;
+	__u32 val;
+	int err;
+
+	len = sizeof(val);
+	err = getsockopt(fd, SOL_IP, IP_LOCAL_PORT_RANGE, &val, &len);
+	if (err)
+		return -1;
+
+	*range = val;
+	return 0;
+}
+
+static const char *so_domain_to_str(int domain)
+{
+	switch (domain) {
+	case AF_INET:
+		return "ip4";
+	case AF_INET6:
+		return "ip6";
+	default:
+		__builtin_unreachable();
+	}
+}
+
+static const char *so_type_to_str(int type)
+{
+	switch (type) {
+	case SOCK_STREAM:
+		return "tcp";
+	case SOCK_DGRAM:
+		return "udp";
+	default:
+		__builtin_unreachable();
+	}
+}
+
+const struct socket_kind {
+	int domain;
+	int type;
+} socket_matrix[] = {
+	{ AF_INET, SOCK_STREAM },
+	{ AF_INET, SOCK_DGRAM },
+	{ AF_INET6, SOCK_STREAM },
+	{ AF_INET6, SOCK_DGRAM },
+};
+
+#define for_each_socket_kind(sk) \
+	for ((sk) = socket_matrix; \
+	     (sk) < socket_matrix + ARRAY_SIZE(socket_matrix) && log_socket_kind((sk)); \
+	     (sk)++)
+
+#define log_socket_kind(sk)			       \
+	({					       \
+		TH_LOG("%s/%s",			       \
+		       so_domain_to_str((sk)->domain), \
+		       so_type_to_str((sk)->type));    \
+		true;				       \
+	})
+
+TEST(invalid_option_value)
+{
+	const struct socket_kind *sk;
+
+	for_each_socket_kind(sk) {
+		__u16 val16;
+		__u32 val32;
+		__u64 val64;
+		int fd, err;
+
+		fd = socket(sk->domain, sk->type, 0);
+		ASSERT_GE(fd, 0) TH_LOG("socket failed");
+
+		/* Too few bytes */
+		val16 = 40000;
+		err = setsockopt(fd, SOL_IP, IP_LOCAL_PORT_RANGE, &val16, sizeof(val16));
+		EXPECT_TRUE(err) TH_LOG("expected setsockopt(IP_LOCAL_PORT_RANGE) to fail");
+		EXPECT_EQ(errno, EINVAL);
+
+		/* Empty range: low port > high port */
+		val32 = pack_port_range(40222, 40111);
+		err = setsockopt(fd, SOL_IP, IP_LOCAL_PORT_RANGE, &val32, sizeof(val32));
+		EXPECT_TRUE(err) TH_LOG("expected setsockopt(IP_LOCAL_PORT_RANGE) to fail");
+		EXPECT_EQ(errno, EINVAL);
+
+		/* Too many bytes */
+		val64 = pack_port_range(40333, 40444);
+		err = setsockopt(fd, SOL_IP, IP_LOCAL_PORT_RANGE, &val64, sizeof(val64));
+		EXPECT_TRUE(err) TH_LOG("expected setsockopt(IP_LOCAL_PORT_RANGE) to fail");
+		EXPECT_EQ(errno, EINVAL);
+
+		err = close(fd);
+		ASSERT_TRUE(!err) TH_LOG("close failed");
+	}
+}
+
+TEST(sock_port_range_out_of_netns_range)
+{
+	const struct socket_kind *sk;
+
+	for_each_socket_kind(sk) {
+		int fd, err, port;
+		__u32 range;
+
+		fd = socket(sk->domain, sk->type, 0);
+		ASSERT_GE(fd, 0) TH_LOG("socket failed");
+
+		range = pack_port_range(50000, 50111);
+		err = setsockopt(fd, SOL_IP, IP_LOCAL_PORT_RANGE, &range, sizeof(range));
+		ASSERT_TRUE(!err) TH_LOG("setsockopt(IP_LOCAL_PORT_RANGE) failed");
+
+		err = bind_to_loopback_any_port(fd);
+		ASSERT_TRUE(!err) TH_LOG("bind failed");
+
+		/* Check that socket port range outside of ephemeral range is ignored */
+		port = get_sock_port(fd);
+		ASSERT_GE(port, 40000) TH_LOG("expected port within ephemeral range");
+		ASSERT_LE(port, 49999) TH_LOG("expected port within ephemeral range");
+
+		err = close(fd);
+		ASSERT_TRUE(!err) TH_LOG("close failed");
+	}
+}
+
+TEST(single_port_range)
+{
+	const struct socket_kind *sk;
+
+	for_each_socket_kind(sk) {
+		struct test {
+			__u16 range_lo;
+			__u16 range_hi;
+			__u16 expected;
+		} tests[] = {
+			/* single port range within ephemeral range */
+			{ 45000, 45000, 45000 },
+			/* first port in the ephemeral range (clamp from above) */
+			{ 0, 40000, 40000 },
+			/* last port in the ephemeral range (clamp from below)  */
+			{ 49999, 0, 49999 },
+		};
+		struct test *t;
+
+		for (t = tests; t < tests + ARRAY_SIZE(tests); t++) {
+			int fd, err, port;
+			__u32 range;
+
+			TH_LOG("lo %5hu, hi %5hu, expected %5hu",
+			       t->range_lo, t->range_hi, t->expected);
+
+			fd = socket(sk->domain, sk->type, 0);
+			ASSERT_GE(fd, 0) TH_LOG("socket failed");
+
+			range = pack_port_range(t->range_lo, t->range_hi);
+			err = setsockopt(fd, SOL_IP, IP_LOCAL_PORT_RANGE, &range, sizeof(range));
+			ASSERT_TRUE(!err) TH_LOG("setsockopt(IP_LOCAL_PORT_RANGE) failed");
+
+			err = bind_to_loopback_any_port(fd);
+			ASSERT_TRUE(!err) TH_LOG("bind failed");
+
+			port = get_sock_port(fd);
+			ASSERT_EQ(port, t->expected) TH_LOG("unexpected local port");
+
+			err = close(fd);
+			ASSERT_TRUE(!err) TH_LOG("close failed");
+		}
+	}
+}
+
+TEST(exhaust_8_port_range)
+{
+	const struct socket_kind *sk;
+
+	for_each_socket_kind(sk) {
+		__u8 port_set = 0;
+		int i, fd, err;
+		__u32 range;
+		__u16 port;
+		int fds[8];
+
+		for (i = 0; i < ARRAY_SIZE(fds); i++) {
+			fd = socket(sk->domain, sk->type, 0);
+			ASSERT_GE(fd, 0) TH_LOG("socket failed");
+
+			range = pack_port_range(40000, 40007);
+			err = setsockopt(fd, SOL_IP, IP_LOCAL_PORT_RANGE, &range, sizeof(range));
+			ASSERT_TRUE(!err) TH_LOG("setsockopt(IP_LOCAL_PORT_RANGE) failed");
+
+			err = bind_to_loopback_any_port(fd);
+			ASSERT_TRUE(!err) TH_LOG("bind failed");
+
+			port = get_sock_port(fd);
+			ASSERT_GE(port, 40000) TH_LOG("expected port within sockopt range");
+			ASSERT_LE(port, 40007) TH_LOG("expected port within sockopt range");
+
+			port_set |= 1 << (port - 40000);
+			fds[i] = fd;
+		}
+
+		/* Check that all every port from the test range is in use */
+		ASSERT_EQ(port_set, 0xff) TH_LOG("expected all ports to be busy");
+
+		/* Check that bind() fails because the whole range is busy */
+		fd = socket(sk->domain, sk->type, 0);
+		ASSERT_GE(fd, 0) TH_LOG("socket failed");
+
+		range = pack_port_range(40000, 40007);
+		err = setsockopt(fd, SOL_IP, IP_LOCAL_PORT_RANGE, &range, sizeof(range));
+		ASSERT_TRUE(!err) TH_LOG("setsockopt(IP_LOCAL_PORT_RANGE) failed");
+
+		err = bind_to_loopback_any_port(fd);
+		ASSERT_TRUE(err) TH_LOG("expected bind to fail");
+		ASSERT_EQ(errno, EADDRINUSE);
+
+		err = close(fd);
+		ASSERT_TRUE(!err) TH_LOG("close failed");
+
+		for (i = 0; i < ARRAY_SIZE(fds); i++) {
+			err = close(fds[i]);
+			ASSERT_TRUE(!err) TH_LOG("close failed");
+		}
+	}
+}
+
+TEST(late_bind)
+{
+	const struct socket_kind *sk;
+
+	for_each_socket_kind(sk) {
+		union {
+			struct sockaddr sa;
+			struct sockaddr_in v4;
+			struct sockaddr_in6 v6;
+		} addr;
+		socklen_t addr_len;
+		const int one = 1;
+		int fd, err;
+		__u32 range;
+		__u16 port;
+
+		fd = socket(sk->domain, sk->type, 0);
+		ASSERT_GE(fd, 0) TH_LOG("socket failed");
+
+		range = pack_port_range(40100, 40199);
+		err = setsockopt(fd, SOL_IP, IP_LOCAL_PORT_RANGE, &range, sizeof(range));
+		ASSERT_TRUE(!err) TH_LOG("setsockopt(IP_LOCAL_PORT_RANGE) failed");
+
+		err = setsockopt(fd, SOL_IP, IP_BIND_ADDRESS_NO_PORT, &one, sizeof(one));
+		ASSERT_TRUE(!err) TH_LOG("setsockopt(IP_BIND_ADDRESS_NO_PORT) failed");
+
+		err = bind_to_loopback_any_port(fd);
+		ASSERT_TRUE(!err) TH_LOG("bind failed");
+
+		port = get_sock_port(fd);
+		ASSERT_EQ(port, 0) TH_LOG("getsockname failed");
+
+		/* Invalid destination */
+		memset(&addr, 0, sizeof(addr));
+		switch (sk->domain) {
+		case AF_INET:
+			addr.v4.sin_family = AF_INET;
+			addr.v4.sin_port = htons(0);
+			addr.v4.sin_addr.s_addr = htonl(INADDR_ANY);
+			addr_len = sizeof(addr.v4);
+			break;
+		case AF_INET6:
+			addr.v6.sin6_family = AF_INET6;
+			addr.v6.sin6_port = htons(0);
+			addr.v6.sin6_addr = in6addr_any;
+			addr_len = sizeof(addr.v6);
+			break;
+		default:
+			ASSERT_TRUE(false) TH_LOG("unsupported socket domain");
+		}
+
+		/* connect() doesn't need to succeed for late bind to happen */
+		connect(fd, &addr.sa, addr_len);
+
+		port = get_sock_port(fd);
+		ASSERT_GE(port, 40100);
+		ASSERT_LE(port, 40199);
+
+		err = close(fd);
+		ASSERT_TRUE(!err) TH_LOG("close failed");
+	}
+}
+
+TEST(get_port_range)
+{
+	const struct socket_kind *sk;
+
+	for_each_socket_kind(sk) {
+		__u16 lo, hi;
+		__u32 range;
+		int fd, err;
+
+		fd = socket(sk->domain, sk->type, 0);
+		ASSERT_GE(fd, 0) TH_LOG("socket failed");
+
+		/* Get range before it will be set */
+		err = get_ip_local_port_range(fd, &range);
+		ASSERT_TRUE(!err) TH_LOG("getsockopt(IP_LOCAL_PORT_RANGE) failed");
+
+		unpack_port_range(range, &lo, &hi);
+		ASSERT_EQ(lo, 0) TH_LOG("unexpected low port");
+		ASSERT_EQ(hi, 0) TH_LOG("unexpected high port");
+
+		range = pack_port_range(12345, 54321);
+		err = setsockopt(fd, SOL_IP, IP_LOCAL_PORT_RANGE, &range, sizeof(range));
+		ASSERT_TRUE(!err) TH_LOG("setsockopt(IP_LOCAL_PORT_RANGE) failed");
+
+		/* Get range after it has been set */
+		err = get_ip_local_port_range(fd, &range);
+		ASSERT_TRUE(!err) TH_LOG("getsockopt(IP_LOCAL_PORT_RANGE) failed");
+
+		unpack_port_range(range, &lo, &hi);
+		ASSERT_EQ(lo, 12345) TH_LOG("unexpected low port");
+		ASSERT_EQ(hi, 54321) TH_LOG("unexpected high port");
+
+		/* Unset the port range  */
+		range = pack_port_range(0, 0);
+		err = setsockopt(fd, SOL_IP, IP_LOCAL_PORT_RANGE, &range, sizeof(range));
+		ASSERT_TRUE(!err) TH_LOG("setsockopt(IP_LOCAL_PORT_RANGE) failed");
+
+		/* Get range after it has been unset */
+		err = get_ip_local_port_range(fd, &range);
+		ASSERT_TRUE(!err) TH_LOG("getsockopt(IP_LOCAL_PORT_RANGE) failed");
+
+		unpack_port_range(range, &lo, &hi);
+		ASSERT_EQ(lo, 0) TH_LOG("unexpected low port");
+		ASSERT_EQ(hi, 0) TH_LOG("unexpected high port");
+
+		err = close(fd);
+		ASSERT_TRUE(!err) TH_LOG("close failed");
+	}
+}
+
+TEST_HARNESS_MAIN
diff --git a/tools/testing/selftests/net/ip_local_port_range.sh b/tools/testing/selftests/net/ip_local_port_range.sh
new file mode 100755
index 000000000000..6c6ad346eaa0
--- /dev/null
+++ b/tools/testing/selftests/net/ip_local_port_range.sh
@@ -0,0 +1,5 @@
+#!/bin/sh
+# SPDX-License-Identifier: GPL-2.0
+
+./in_netns.sh \
+  sh -c 'sysctl -q -w net.ipv4.ip_local_port_range="40000 49999" && ./ip_local_port_range'

-- 
2.38.1

Re: [PATCH net-next 1/2] inet: Add IP_LOCAL_PORT_RANGE socket option

[Kuniyuki Iwashima]

From:   Jakub Sitnicki <jakub@cloudflare.com>
Date:   Fri,  6 Jan 2023 11:37:37 +0100
> Users who want to share a single public IP address for outgoing connections
> between several hosts traditionally reach for SNAT. However, SNAT requires
> state keeping on the node(s) performing the NAT.
> 
> A stateless alternative exists, where a single IP address used for egress
> can be shared between several hosts by partitioning the available ephemeral
> port range. In such a setup:
> 
> 1. Each host gets assigned a disjoint range of ephemeral ports.
> 2. Applications open connections from the host-assigned port range.
> 3. Return traffic gets routed to the host based on both, the destination IP
>    and the destination port.
> 
> An application which wants to open an outgoing connection (connect) from a
> given port range today can choose between two solutions:
> 
> 1. Manually pick the source port by bind()'ing to it before connect()'ing
>    the socket.
> 
>    This approach has a couple of downsides:
> 
>    a) Search for a free port has to be implemented in the user-space. If
>       the chosen 4-tuple happens to be busy, the application needs to retry
>       from a different local port number.
> 
>       Detecting if 4-tuple is busy can be either easy (TCP) or hard
>       (UDP). In TCP case, the application simply has to check if connect()
>       returned an error (EADDRNOTAVAIL). That is assuming that the local
>       port sharing was enabled (REUSEADDR) by all the sockets.
> 
>         # Assume desired local port range is 60_000-60_511
>         s = socket(AF_INET, SOCK_STREAM)
>         s.setsockopt(SOL_SOCKET, SO_REUSEADDR, 1)
>         s.bind(("192.0.2.1", 60_000))
>         s.connect(("1.1.1.1", 53))
>         # Fails only if 192.0.2.1:60000 -> 1.1.1.1:53 is busy
>         # Application must retry with another local port
> 
>       In case of UDP, the network stack allows binding more than one socket
>       to the same 4-tuple, when local port sharing is enabled
>       (REUSEADDR). Hence detecting the conflict is much harder and involves
>       querying sock_diag and toggling the REUSEADDR flag [1].
> 
>    b) For TCP, bind()-ing to a port within the ephemeral port range means
>       that no connecting sockets, that is those which leave it to the
>       network stack to find a free local port at connect() time, can use
>       the this port.
> 
>       IOW, the bind hash bucket tb->fastreuse will be 0 or 1, and the port
>       will be skipped during the free port search at connect() time.
> 
> 2. Isolate the app in a dedicated netns and use the use the per-netns
>    ip_local_port_range sysctl to adjust the ephemeral port range bounds.
> 
>    The per-netns setting affects all sockets, so this approach can be used
>    only if:
> 
>    - there is just one egress IP address, or
>    - the desired egress port range is the same for all egress IP addresses
>      used by the application.
> 
>    For TCP, this approach avoids the downsides of (1). Free port search and
>    4-tuple conflict detection is done by the network stack:
> 
>      system("sysctl -w net.ipv4.ip_local_port_range='60000 60511'")
> 
>      s = socket(AF_INET, SOCK_STREAM)
>      s.setsockopt(SOL_IP, IP_BIND_ADDRESS_NO_PORT, 1)
>      s.bind(("192.0.2.1", 0))
>      s.connect(("1.1.1.1", 53))
>      # Fails if all 4-tuples 192.0.2.1:60000-60511 -> 1.1.1.1:53 are busy
> 
>   For UDP this approach has limited applicability. Setting the
>   IP_BIND_ADDRESS_NO_PORT socket option does not result in local source
>   port being shared with other connected UDP sockets.
> 
>   Hence relying on the network stack to find a free source port, limits the
>   number of outgoing UDP flows from a single IP address down to the number
>   of available ephemeral ports.
> 
> To put it another way, partitioning the ephemeral port range between hosts
> using the existing Linux networking API is cumbersome.
> 
> To address this use case, add a new socket option at the SOL_IP level,
> named IP_LOCAL_PORT_RANGE. The new option can be used to clamp down the
> ephemeral port range for each socket individually.
> 
> The option can be used only to narrow down the per-netns local port
> range. If the per-socket range lies outside of the per-netns range, the
> latter takes precedence.
> 
> UAPI-wise, the low and high range bounds are passed to the kernel as a pair
> of u16 values packed into a u32. This avoids pointer passing.
> 
>   PORT_LO = 40_000
>   PORT_HI = 40_511
> 
>   s = socket(AF_INET, SOCK_STREAM)
>   v = struct.pack("I", PORT_LO | (PORT_HI << 16))
>   s.setsockopt(SOL_IP, IP_LOCAL_PORT_RANGE, v)
>   s.bind(("127.0.0.1", 0))
>   s.getsockname()
>   # Local address between ("127.0.0.1", 40_000) and ("127.0.0.1", 40_511),
>   # if there is a free port. EADDRINUSE otherwise.
> 
> [1] https://github.com/cloudflare/cloudflare-blog/blob/232b432c1d57/2022-02-connectx/connectx.py#L116
> 
> Reviewed-by: Marek Majkowski <marek@cloudflare.com>
> Signed-off-by: Jakub Sitnicki <jakub@cloudflare.com>
> ---
>  include/net/inet_sock.h         |  4 ++++
>  include/net/ip.h                |  3 ++-
>  include/uapi/linux/in.h         |  1 +
>  net/ipv4/inet_connection_sock.c | 22 ++++++++++++++++++++--
>  net/ipv4/inet_hashtables.c      |  2 +-
>  net/ipv4/ip_sockglue.c          | 18 ++++++++++++++++++
>  net/ipv4/udp.c                  |  2 +-
>  7 files changed, 47 insertions(+), 5 deletions(-)
> 
> diff --git a/include/net/inet_sock.h b/include/net/inet_sock.h
> index bf5654ce711e..51857117ac09 100644
> --- a/include/net/inet_sock.h
> +++ b/include/net/inet_sock.h
> @@ -249,6 +249,10 @@ struct inet_sock {
>  	__be32			mc_addr;
>  	struct ip_mc_socklist __rcu	*mc_list;
>  	struct inet_cork_full	cork;
> +	struct {
> +		__u16 lo;
> +		__u16 hi;
> +	}			local_port_range;
>  };
>  
>  #define IPCORK_OPT	1	/* ip-options has been held in ipcork.opt */
> diff --git a/include/net/ip.h b/include/net/ip.h
> index 144bdfbb25af..c3fffaa92d6e 100644
> --- a/include/net/ip.h
> +++ b/include/net/ip.h
> @@ -340,7 +340,8 @@ static inline u64 snmp_fold_field64(void __percpu *mib, int offt, size_t syncp_o
>  	} \
>  }
>  
> -void inet_get_local_port_range(struct net *net, int *low, int *high);
> +void inet_get_local_port_range(const struct net *net, int *low, int *high);
> +void inet_sk_get_local_port_range(const struct sock *sk, int *low, int *high);
>  
>  #ifdef CONFIG_SYSCTL
>  static inline bool inet_is_local_reserved_port(struct net *net, unsigned short port)
> diff --git a/include/uapi/linux/in.h b/include/uapi/linux/in.h
> index 07a4cb149305..4b7f2df66b99 100644
> --- a/include/uapi/linux/in.h
> +++ b/include/uapi/linux/in.h
> @@ -162,6 +162,7 @@ struct in_addr {
>  #define MCAST_MSFILTER			48
>  #define IP_MULTICAST_ALL		49
>  #define IP_UNICAST_IF			50
> +#define IP_LOCAL_PORT_RANGE		51
>  
>  #define MCAST_EXCLUDE	0
>  #define MCAST_INCLUDE	1
> diff --git a/net/ipv4/inet_connection_sock.c b/net/ipv4/inet_connection_sock.c
> index d1f837579398..ec40c9ee753c 100644
> --- a/net/ipv4/inet_connection_sock.c
> +++ b/net/ipv4/inet_connection_sock.c
> @@ -117,7 +117,7 @@ bool inet_rcv_saddr_any(const struct sock *sk)
>  	return !sk->sk_rcv_saddr;
>  }
>  
> -void inet_get_local_port_range(struct net *net, int *low, int *high)
> +void inet_get_local_port_range(const struct net *net, int *low, int *high)
>  {
>  	unsigned int seq;
>  
> @@ -130,6 +130,24 @@ void inet_get_local_port_range(struct net *net, int *low, int *high)
>  }
>  EXPORT_SYMBOL(inet_get_local_port_range);
>  
> +void inet_sk_get_local_port_range(const struct sock *sk, int *low, int *high)
> +{
> +	const struct inet_sock *inet = inet_sk(sk);
> +	const struct net *net = sock_net(sk);
> +	int lo, hi;
> +
> +	inet_get_local_port_range(net, &lo, &hi);
> +
> +	if (unlikely(inet->local_port_range.lo))
> +		lo = clamp_val(inet->local_port_range.lo, lo, hi);
> +	if (unlikely(inet->local_port_range.hi))
> +		hi = clamp_val(inet->local_port_range.hi, lo, hi);

If both vals are outside of the global range, the new range is clamped
to (netns-lo, netns-lo) or (netnsl-hi, netns-hi).

    .lo   .hi     lo                 hi     .lo    .hi
     |-----|       |-----------------|       |------|

It seems the description in the man page and changelog is not correct.


> +
> +	*low = lo;
> +	*high = hi;
> +}
> +EXPORT_SYMBOL(inet_sk_get_local_port_range);
> +
>  static bool inet_use_bhash2_on_bind(const struct sock *sk)
>  {
>  #if IS_ENABLED(CONFIG_IPV6)
> @@ -316,7 +334,7 @@ inet_csk_find_open_port(const struct sock *sk, struct inet_bind_bucket **tb_ret,
>  ports_exhausted:
>  	attempt_half = (sk->sk_reuse == SK_CAN_REUSE) ? 1 : 0;
>  other_half_scan:
> -	inet_get_local_port_range(net, &low, &high);
> +	inet_sk_get_local_port_range(sk, &low, &high);
>  	high++; /* [32768, 60999] -> [32768, 61000[ */
>  	if (high - low < 4)
>  		attempt_half = 0;
> diff --git a/net/ipv4/inet_hashtables.c b/net/ipv4/inet_hashtables.c
> index 24a38b56fab9..22ee5f23ee5b 100644
> --- a/net/ipv4/inet_hashtables.c
> +++ b/net/ipv4/inet_hashtables.c
> @@ -1013,7 +1013,7 @@ int __inet_hash_connect(struct inet_timewait_death_row *death_row,
>  
>  	l3mdev = inet_sk_bound_l3mdev(sk);
>  
> -	inet_get_local_port_range(net, &low, &high);
> +	inet_sk_get_local_port_range(sk, &low, &high);
>  	high++; /* [32768, 60999] -> [32768, 61000[ */
>  	remaining = high - low;
>  	if (likely(remaining > 1))
> diff --git a/net/ipv4/ip_sockglue.c b/net/ipv4/ip_sockglue.c
> index 9f92ae35bb01..b511ff0adc0a 100644
> --- a/net/ipv4/ip_sockglue.c
> +++ b/net/ipv4/ip_sockglue.c
> @@ -923,6 +923,7 @@ int do_ip_setsockopt(struct sock *sk, int level, int optname,
>  	case IP_CHECKSUM:
>  	case IP_RECVFRAGSIZE:
>  	case IP_RECVERR_RFC4884:
> +	case IP_LOCAL_PORT_RANGE:
>  		if (optlen >= sizeof(int)) {
>  			if (copy_from_sockptr(&val, optval, sizeof(val)))
>  				return -EFAULT;
> @@ -1365,6 +1366,20 @@ int do_ip_setsockopt(struct sock *sk, int level, int optname,
>  		WRITE_ONCE(inet->min_ttl, val);
>  		break;
>  
> +	case IP_LOCAL_PORT_RANGE:
> +	{
> +		const __u16 lo = val;
> +		const __u16 hi = val >> 16;
> +
> +		if (optlen != sizeof(__u32))
> +			goto e_inval;
> +		if (lo != 0 && hi != 0 && lo > hi)

Should (0, 0) be EINVAL as it has no effect ?

                if ((!lo && !hi) || (lo && hi && lo > hi))
                        goto e_inval;

Thank you,
Kuniyuki


> +			goto e_inval;
> +
> +		inet->local_port_range.lo = lo;
> +		inet->local_port_range.hi = hi;
> +		break;
> +	}
>  	default:
>  		err = -ENOPROTOOPT;
>  		break;
> @@ -1743,6 +1758,9 @@ int do_ip_getsockopt(struct sock *sk, int level, int optname,
>  	case IP_MINTTL:
>  		val = inet->min_ttl;
>  		break;
> +	case IP_LOCAL_PORT_RANGE:
> +		val = inet->local_port_range.hi << 16 | inet->local_port_range.lo;
> +		break;
>  	default:
>  		sockopt_release_sock(sk);
>  		return -ENOPROTOOPT;
> diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c
> index 9592fe3e444a..c605d171eb2d 100644
> --- a/net/ipv4/udp.c
> +++ b/net/ipv4/udp.c
> @@ -248,7 +248,7 @@ int udp_lib_get_port(struct sock *sk, unsigned short snum,
>  		int low, high, remaining;
>  		unsigned int rand;
>  
> -		inet_get_local_port_range(net, &low, &high);
> +		inet_sk_get_local_port_range(sk, &low, &high);
>  		remaining = (high - low) + 1;
>  
>  		rand = get_random_u32();
> 
> -- 
> 2.38.1

Re: [PATCH net-next 1/2] inet: Add IP_LOCAL_PORT_RANGE socket option

[Jakub Sitnicki]

On Sat, Jan 07, 2023 at 02:16 AM +09, Kuniyuki Iwashima wrote:
> From:   Jakub Sitnicki <jakub@cloudflare.com>
> Date:   Fri,  6 Jan 2023 11:37:37 +0100
>> Users who want to share a single public IP address for outgoing connections
>> between several hosts traditionally reach for SNAT. However, SNAT requires
>> state keeping on the node(s) performing the NAT.
>> 
>> A stateless alternative exists, where a single IP address used for egress
>> can be shared between several hosts by partitioning the available ephemeral
>> port range. In such a setup:
>> 
>> 1. Each host gets assigned a disjoint range of ephemeral ports.
>> 2. Applications open connections from the host-assigned port range.
>> 3. Return traffic gets routed to the host based on both, the destination IP
>>    and the destination port.
>> 
>> An application which wants to open an outgoing connection (connect) from a
>> given port range today can choose between two solutions:
>> 
>> 1. Manually pick the source port by bind()'ing to it before connect()'ing
>>    the socket.
>> 
>>    This approach has a couple of downsides:
>> 
>>    a) Search for a free port has to be implemented in the user-space. If
>>       the chosen 4-tuple happens to be busy, the application needs to retry
>>       from a different local port number.
>> 
>>       Detecting if 4-tuple is busy can be either easy (TCP) or hard
>>       (UDP). In TCP case, the application simply has to check if connect()
>>       returned an error (EADDRNOTAVAIL). That is assuming that the local
>>       port sharing was enabled (REUSEADDR) by all the sockets.
>> 
>>         # Assume desired local port range is 60_000-60_511
>>         s = socket(AF_INET, SOCK_STREAM)
>>         s.setsockopt(SOL_SOCKET, SO_REUSEADDR, 1)
>>         s.bind(("192.0.2.1", 60_000))
>>         s.connect(("1.1.1.1", 53))
>>         # Fails only if 192.0.2.1:60000 -> 1.1.1.1:53 is busy
>>         # Application must retry with another local port
>> 
>>       In case of UDP, the network stack allows binding more than one socket
>>       to the same 4-tuple, when local port sharing is enabled
>>       (REUSEADDR). Hence detecting the conflict is much harder and involves
>>       querying sock_diag and toggling the REUSEADDR flag [1].
>> 
>>    b) For TCP, bind()-ing to a port within the ephemeral port range means
>>       that no connecting sockets, that is those which leave it to the
>>       network stack to find a free local port at connect() time, can use
>>       the this port.
>> 
>>       IOW, the bind hash bucket tb->fastreuse will be 0 or 1, and the port
>>       will be skipped during the free port search at connect() time.
>> 
>> 2. Isolate the app in a dedicated netns and use the use the per-netns
>>    ip_local_port_range sysctl to adjust the ephemeral port range bounds.
>> 
>>    The per-netns setting affects all sockets, so this approach can be used
>>    only if:
>> 
>>    - there is just one egress IP address, or
>>    - the desired egress port range is the same for all egress IP addresses
>>      used by the application.
>> 
>>    For TCP, this approach avoids the downsides of (1). Free port search and
>>    4-tuple conflict detection is done by the network stack:
>> 
>>      system("sysctl -w net.ipv4.ip_local_port_range='60000 60511'")
>> 
>>      s = socket(AF_INET, SOCK_STREAM)
>>      s.setsockopt(SOL_IP, IP_BIND_ADDRESS_NO_PORT, 1)
>>      s.bind(("192.0.2.1", 0))
>>      s.connect(("1.1.1.1", 53))
>>      # Fails if all 4-tuples 192.0.2.1:60000-60511 -> 1.1.1.1:53 are busy
>> 
>>   For UDP this approach has limited applicability. Setting the
>>   IP_BIND_ADDRESS_NO_PORT socket option does not result in local source
>>   port being shared with other connected UDP sockets.
>> 
>>   Hence relying on the network stack to find a free source port, limits the
>>   number of outgoing UDP flows from a single IP address down to the number
>>   of available ephemeral ports.
>> 
>> To put it another way, partitioning the ephemeral port range between hosts
>> using the existing Linux networking API is cumbersome.
>> 
>> To address this use case, add a new socket option at the SOL_IP level,
>> named IP_LOCAL_PORT_RANGE. The new option can be used to clamp down the
>> ephemeral port range for each socket individually.
>> 
>> The option can be used only to narrow down the per-netns local port
>> range. If the per-socket range lies outside of the per-netns range, the
>> latter takes precedence.
>> 
>> UAPI-wise, the low and high range bounds are passed to the kernel as a pair
>> of u16 values packed into a u32. This avoids pointer passing.
>> 
>>   PORT_LO = 40_000
>>   PORT_HI = 40_511
>> 
>>   s = socket(AF_INET, SOCK_STREAM)
>>   v = struct.pack("I", PORT_LO | (PORT_HI << 16))
>>   s.setsockopt(SOL_IP, IP_LOCAL_PORT_RANGE, v)
>>   s.bind(("127.0.0.1", 0))
>>   s.getsockname()
>>   # Local address between ("127.0.0.1", 40_000) and ("127.0.0.1", 40_511),
>>   # if there is a free port. EADDRINUSE otherwise.
>> 
>> [1] https://github.com/cloudflare/cloudflare-blog/blob/232b432c1d57/2022-02-connectx/connectx.py#L116
>> 
>> Reviewed-by: Marek Majkowski <marek@cloudflare.com>
>> Signed-off-by: Jakub Sitnicki <jakub@cloudflare.com>
>> ---

[...]

>> --- a/net/ipv4/inet_connection_sock.c
>> +++ b/net/ipv4/inet_connection_sock.c
>> @@ -117,7 +117,7 @@ bool inet_rcv_saddr_any(const struct sock *sk)
>>  	return !sk->sk_rcv_saddr;
>>  }
>>  
>> -void inet_get_local_port_range(struct net *net, int *low, int *high)
>> +void inet_get_local_port_range(const struct net *net, int *low, int *high)
>>  {
>>  	unsigned int seq;
>>  
>> @@ -130,6 +130,24 @@ void inet_get_local_port_range(struct net *net, int *low, int *high)
>>  }
>>  EXPORT_SYMBOL(inet_get_local_port_range);
>>  
>> +void inet_sk_get_local_port_range(const struct sock *sk, int *low, int *high)
>> +{
>> +	const struct inet_sock *inet = inet_sk(sk);
>> +	const struct net *net = sock_net(sk);
>> +	int lo, hi;
>> +
>> +	inet_get_local_port_range(net, &lo, &hi);
>> +
>> +	if (unlikely(inet->local_port_range.lo))
>> +		lo = clamp_val(inet->local_port_range.lo, lo, hi);
>> +	if (unlikely(inet->local_port_range.hi))
>> +		hi = clamp_val(inet->local_port_range.hi, lo, hi);
>
> If both vals are outside of the global range, the new range is clamped
> to (netns-lo, netns-lo) or (netnsl-hi, netns-hi).
>
>     .lo   .hi     lo                 hi     .lo    .hi
>      |-----|       |-----------------|       |------|
>
> It seems the description in the man page and changelog is not correct.

This is a bug. I overlooked this corner case.
Thank you for pointing it out.
Will fix and add test coverage in v2.

[...]

>> --- a/net/ipv4/ip_sockglue.c
>> +++ b/net/ipv4/ip_sockglue.c
>> @@ -923,6 +923,7 @@ int do_ip_setsockopt(struct sock *sk, int level, int optname,
>>  	case IP_CHECKSUM:
>>  	case IP_RECVFRAGSIZE:
>>  	case IP_RECVERR_RFC4884:
>> +	case IP_LOCAL_PORT_RANGE:
>>  		if (optlen >= sizeof(int)) {
>>  			if (copy_from_sockptr(&val, optval, sizeof(val)))
>>  				return -EFAULT;
>> @@ -1365,6 +1366,20 @@ int do_ip_setsockopt(struct sock *sk, int level, int optname,
>>  		WRITE_ONCE(inet->min_ttl, val);
>>  		break;
>>  
>> +	case IP_LOCAL_PORT_RANGE:
>> +	{
>> +		const __u16 lo = val;
>> +		const __u16 hi = val >> 16;
>> +
>> +		if (optlen != sizeof(__u32))
>> +			goto e_inval;
>> +		if (lo != 0 && hi != 0 && lo > hi)
>
> Should (0, 0) be EINVAL as it has no effect ?
>
>                 if ((!lo && !hi) || (lo && hi && lo > hi))
>                         goto e_inval;

User can pass (0, 0) to unset the setting. This is intentional.
The `get_port_range` test in the following patch covers it.

Thank you for feedback,
Jakub

Re: [PATCH net-next 1/2] inet: Add IP_LOCAL_PORT_RANGE socket option

[Kuniyuki Iwashima]

From:   Jakub Sitnicki <jakub@cloudflare.com>
Date:   Mon, 09 Jan 2023 11:11:32 +0100
> On Sat, Jan 07, 2023 at 02:16 AM +09, Kuniyuki Iwashima wrote:
> > From:   Jakub Sitnicki <jakub@cloudflare.com>
> > Date:   Fri,  6 Jan 2023 11:37:37 +0100
> >> Users who want to share a single public IP address for outgoing connections
> >> between several hosts traditionally reach for SNAT. However, SNAT requires
> >> state keeping on the node(s) performing the NAT.
> >> 
> >> A stateless alternative exists, where a single IP address used for egress
> >> can be shared between several hosts by partitioning the available ephemeral
> >> port range. In such a setup:
> >> 
> >> 1. Each host gets assigned a disjoint range of ephemeral ports.
> >> 2. Applications open connections from the host-assigned port range.
> >> 3. Return traffic gets routed to the host based on both, the destination IP
> >>    and the destination port.
> >> 
> >> An application which wants to open an outgoing connection (connect) from a
> >> given port range today can choose between two solutions:
> >> 
> >> 1. Manually pick the source port by bind()'ing to it before connect()'ing
> >>    the socket.
> >> 
> >>    This approach has a couple of downsides:
> >> 
> >>    a) Search for a free port has to be implemented in the user-space. If
> >>       the chosen 4-tuple happens to be busy, the application needs to retry
> >>       from a different local port number.
> >> 
> >>       Detecting if 4-tuple is busy can be either easy (TCP) or hard
> >>       (UDP). In TCP case, the application simply has to check if connect()
> >>       returned an error (EADDRNOTAVAIL). That is assuming that the local
> >>       port sharing was enabled (REUSEADDR) by all the sockets.
> >> 
> >>         # Assume desired local port range is 60_000-60_511
> >>         s = socket(AF_INET, SOCK_STREAM)
> >>         s.setsockopt(SOL_SOCKET, SO_REUSEADDR, 1)
> >>         s.bind(("192.0.2.1", 60_000))
> >>         s.connect(("1.1.1.1", 53))
> >>         # Fails only if 192.0.2.1:60000 -> 1.1.1.1:53 is busy
> >>         # Application must retry with another local port
> >> 
> >>       In case of UDP, the network stack allows binding more than one socket
> >>       to the same 4-tuple, when local port sharing is enabled
> >>       (REUSEADDR). Hence detecting the conflict is much harder and involves
> >>       querying sock_diag and toggling the REUSEADDR flag [1].
> >> 
> >>    b) For TCP, bind()-ing to a port within the ephemeral port range means
> >>       that no connecting sockets, that is those which leave it to the
> >>       network stack to find a free local port at connect() time, can use
> >>       the this port.
> >> 
> >>       IOW, the bind hash bucket tb->fastreuse will be 0 or 1, and the port
> >>       will be skipped during the free port search at connect() time.
> >> 
> >> 2. Isolate the app in a dedicated netns and use the use the per-netns
> >>    ip_local_port_range sysctl to adjust the ephemeral port range bounds.
> >> 
> >>    The per-netns setting affects all sockets, so this approach can be used
> >>    only if:
> >> 
> >>    - there is just one egress IP address, or
> >>    - the desired egress port range is the same for all egress IP addresses
> >>      used by the application.
> >> 
> >>    For TCP, this approach avoids the downsides of (1). Free port search and
> >>    4-tuple conflict detection is done by the network stack:
> >> 
> >>      system("sysctl -w net.ipv4.ip_local_port_range='60000 60511'")
> >> 
> >>      s = socket(AF_INET, SOCK_STREAM)
> >>      s.setsockopt(SOL_IP, IP_BIND_ADDRESS_NO_PORT, 1)
> >>      s.bind(("192.0.2.1", 0))
> >>      s.connect(("1.1.1.1", 53))
> >>      # Fails if all 4-tuples 192.0.2.1:60000-60511 -> 1.1.1.1:53 are busy
> >> 
> >>   For UDP this approach has limited applicability. Setting the
> >>   IP_BIND_ADDRESS_NO_PORT socket option does not result in local source
> >>   port being shared with other connected UDP sockets.
> >> 
> >>   Hence relying on the network stack to find a free source port, limits the
> >>   number of outgoing UDP flows from a single IP address down to the number
> >>   of available ephemeral ports.
> >> 
> >> To put it another way, partitioning the ephemeral port range between hosts
> >> using the existing Linux networking API is cumbersome.
> >> 
> >> To address this use case, add a new socket option at the SOL_IP level,
> >> named IP_LOCAL_PORT_RANGE. The new option can be used to clamp down the
> >> ephemeral port range for each socket individually.
> >> 
> >> The option can be used only to narrow down the per-netns local port
> >> range. If the per-socket range lies outside of the per-netns range, the
> >> latter takes precedence.
> >> 
> >> UAPI-wise, the low and high range bounds are passed to the kernel as a pair
> >> of u16 values packed into a u32. This avoids pointer passing.
> >> 
> >>   PORT_LO = 40_000
> >>   PORT_HI = 40_511
> >> 
> >>   s = socket(AF_INET, SOCK_STREAM)
> >>   v = struct.pack("I", PORT_LO | (PORT_HI << 16))
> >>   s.setsockopt(SOL_IP, IP_LOCAL_PORT_RANGE, v)
> >>   s.bind(("127.0.0.1", 0))
> >>   s.getsockname()
> >>   # Local address between ("127.0.0.1", 40_000) and ("127.0.0.1", 40_511),
> >>   # if there is a free port. EADDRINUSE otherwise.
> >> 
> >> [1] https://github.com/cloudflare/cloudflare-blog/blob/232b432c1d57/2022-02-connectx/connectx.py#L116
> >> 
> >> Reviewed-by: Marek Majkowski <marek@cloudflare.com>
> >> Signed-off-by: Jakub Sitnicki <jakub@cloudflare.com>
> >> ---
> 
> [...]
> 
> >> --- a/net/ipv4/inet_connection_sock.c
> >> +++ b/net/ipv4/inet_connection_sock.c
> >> @@ -117,7 +117,7 @@ bool inet_rcv_saddr_any(const struct sock *sk)
> >>  	return !sk->sk_rcv_saddr;
> >>  }
> >>  
> >> -void inet_get_local_port_range(struct net *net, int *low, int *high)
> >> +void inet_get_local_port_range(const struct net *net, int *low, int *high)
> >>  {
> >>  	unsigned int seq;
> >>  
> >> @@ -130,6 +130,24 @@ void inet_get_local_port_range(struct net *net, int *low, int *high)
> >>  }
> >>  EXPORT_SYMBOL(inet_get_local_port_range);
> >>  
> >> +void inet_sk_get_local_port_range(const struct sock *sk, int *low, int *high)
> >> +{
> >> +	const struct inet_sock *inet = inet_sk(sk);
> >> +	const struct net *net = sock_net(sk);
> >> +	int lo, hi;
> >> +
> >> +	inet_get_local_port_range(net, &lo, &hi);
> >> +
> >> +	if (unlikely(inet->local_port_range.lo))
> >> +		lo = clamp_val(inet->local_port_range.lo, lo, hi);
> >> +	if (unlikely(inet->local_port_range.hi))
> >> +		hi = clamp_val(inet->local_port_range.hi, lo, hi);
> >
> > If both vals are outside of the global range, the new range is clamped
> > to (netns-lo, netns-lo) or (netnsl-hi, netns-hi).
> >
> >     .lo   .hi     lo                 hi     .lo    .hi
> >      |-----|       |-----------------|       |------|
> >
> > It seems the description in the man page and changelog is not correct.
> 
> This is a bug. I overlooked this corner case.
> Thank you for pointing it out.
> Will fix and add test coverage in v2.
> 
> [...]
> 
> >> --- a/net/ipv4/ip_sockglue.c
> >> +++ b/net/ipv4/ip_sockglue.c
> >> @@ -923,6 +923,7 @@ int do_ip_setsockopt(struct sock *sk, int level, int optname,
> >>  	case IP_CHECKSUM:
> >>  	case IP_RECVFRAGSIZE:
> >>  	case IP_RECVERR_RFC4884:
> >> +	case IP_LOCAL_PORT_RANGE:
> >>  		if (optlen >= sizeof(int)) {
> >>  			if (copy_from_sockptr(&val, optval, sizeof(val)))
> >>  				return -EFAULT;
> >> @@ -1365,6 +1366,20 @@ int do_ip_setsockopt(struct sock *sk, int level, int optname,
> >>  		WRITE_ONCE(inet->min_ttl, val);
> >>  		break;
> >>  
> >> +	case IP_LOCAL_PORT_RANGE:
> >> +	{
> >> +		const __u16 lo = val;
> >> +		const __u16 hi = val >> 16;
> >> +
> >> +		if (optlen != sizeof(__u32))
> >> +			goto e_inval;
> >> +		if (lo != 0 && hi != 0 && lo > hi)
> >
> > Should (0, 0) be EINVAL as it has no effect ?
> >
> >                 if ((!lo && !hi) || (lo && hi && lo > hi))
> >                         goto e_inval;
> 
> User can pass (0, 0) to unset the setting. This is intentional.
> The `get_port_range` test in the following patch covers it.

That makes sense.
I'll take a look on selftest.
Thanks for explaining!

> 
> Thank you for feedback,
> Jakub