public inbox for passt-dev@passt.top
 help / color / mirror / code / Atom feed
From: Stefano Brivio <sbrivio@redhat.com>
To: passt-dev@passt.top
Cc: Gerwim Feiken <info@gerwim.nl>,
	David Gibson <david@gibson.dropbear.id.au>
Subject: [FOR REFERENCE ONLY] [PATCH] udp: Don't share a preserved source port with other sockets
Date: Sun, 15 Nov 2026 23:38:23 +0100	[thread overview]
Message-ID: <20261115223823.1401886-1-sbrivio@redhat.com> (raw)

From: Gerwim Feiken <info@gerwim.nl>

For a UDP flow initiated from the guest, passt keeps the guest's source port
on the host side (fwd_nat_from_tap()) and binds the flow socket with
SO_REUSEADDR (sock_l4_()). Several passt instances in one network namespace --
one per VM, as a VM manager spawning passt from one daemon does -- can then
each bind the same port towards the same destination, typically the resolver
behind --dns-forward. Their sockets share one identical 4-tuple, and the
kernel delivers each reply to only one of them: the other guest's queries go
unanswered until the flow expires, and the guest owning the receiving socket
gets answers to queries it never sent.

Bind the host side socket of a flow initiated from the tap interface without
SO_REUSEADDR, so that the bind fails if any other socket holds the port, and
fall back to a port chosen by the kernel in that case. udp_flow_new() reads
the actual port back with getsockname(), as it already does for the address.
Listening sockets, and flows initiated from them, keep SO_REUSEADDR.

Signed-off-by: Gerwim Feiken <info@gerwim.nl>
---
This was shared by Gerwim a while ago but I never followed up with it.  Now
that we have https://bugs.passt.top/show_bug.cgi?id=223, I would like to refer
to this patch from that ticket, that's why I'm posting it.

I guess we could solve both issues, the one originally fixed by this patch
and the one raised in https://bugs.passt.top/show_bug.cgi?id=223, using
just this patch, as it implements a fallback for bind() anyway.  But I
didn't review this patch thoroughly yet.

 flow.c     | 11 +++++++----
 flow.h     |  2 +-
 icmp.c     |  2 +-
 pif.c      |  2 +-
 udp_flow.c | 28 ++++++++++++++++++++++++----
 util.c     | 15 ++++++++++-----
 util.h     |  3 ++-
 7 files changed, 46 insertions(+), 17 deletions(-)

diff --git a/flow.c b/flow.c
index 27a24a4..c6ebaf6 100644
--- a/flow.c
+++ b/flow.c
@@ -177,6 +177,7 @@ static void flowside_from_af(struct flowside *side, sa_family_t af,
  * @err:	Filled in with errno if something failed
  * @type:	Socket epoll type
  * @sa:		Socket address
+ * @exclusive:	Don't set SO_REUSEADDR (see sock_l4_())
  */
 struct flowside_sock_args {
 	const struct ctx *c;
@@ -184,6 +185,7 @@ struct flowside_sock_args {
 	int err;
 	enum epoll_type type;
 	const union sockaddr_inany *sa;
+	bool exclusive;
 };
 
 /** flowside_sock_splice() - Create and bind socket for PIF_SPLICE based on flowside
@@ -197,7 +199,7 @@ static int flowside_sock_splice(void *arg)
 
 	ns_enter(a->c);
 
-	a->fd = sock_l4(a->c, a->type, a->sa, NULL);
+	a->fd = sock_l4(a->c, a->type, a->sa, NULL, a->exclusive);
 	a->err = errno;
 
 	return 0;
@@ -208,12 +210,13 @@ static int flowside_sock_splice(void *arg)
  * @type:	Socket epoll type
  * @pif:	Interface for this socket
  * @side:	Flowside to create a socket for
+ * @exclusive:	Don't set SO_REUSEADDR (see sock_l4_())
  *
  * Return: socket fd of protocol @proto bound to our address and port from @side
  *         (if specified).
  */
 int flowside_sock_l4(const struct ctx *c, enum epoll_type type, uint8_t pif,
-		     const struct flowside *side)
+		     const struct flowside *side, bool exclusive)
 {
 	const char *ifname = NULL;
 	union sockaddr_inany sa;
@@ -231,11 +234,11 @@ int flowside_sock_l4(const struct ctx *c, enum epoll_type type, uint8_t pif,
 		else if (sa.sa_family == AF_INET6)
 			ifname = c->ip6.ifname_out;
 
-		return sock_l4(c, type, &sa, ifname);
+		return sock_l4(c, type, &sa, ifname, exclusive);
 
 	case PIF_SPLICE: {
 		struct flowside_sock_args args = {
-			.c = c, .type = type, .sa = &sa,
+			.c = c, .type = type, .sa = &sa, .exclusive = exclusive,
 		};
 		NS_CALL(flowside_sock_splice, &args);
 		errno = args.err;
diff --git a/flow.h b/flow.h
index db02845..3b37ac3 100644
--- a/flow.h
+++ b/flow.h
@@ -173,7 +173,7 @@ static inline bool flowside_eq(const struct flowside *left,
 }
 
 int flowside_sock_l4(const struct ctx *c, enum epoll_type type, uint8_t pif,
-		     const struct flowside *side);
+		     const struct flowside *side, bool exclusive);
 int flowside_connect(const struct ctx *c, int s,
 		     uint8_t pif, const struct flowside *tgt);
 
diff --git a/icmp.c b/icmp.c
index 0fe2366..7499a39 100644
--- a/icmp.c
+++ b/icmp.c
@@ -203,7 +203,7 @@ static struct icmp_ping_flow *icmp_ping_new(const struct ctx *c,
 
 	pingf->seq = -1;
 
-	pingf->sock = flowside_sock_l4(c, EPOLL_TYPE_PING, PIF_HOST, tgt);
+	pingf->sock = flowside_sock_l4(c, EPOLL_TYPE_PING, PIF_HOST, tgt, false);
 
 	if (pingf->sock < 0) {
 		warn("Cannot open \"ping\" socket. You might need to:");
diff --git a/pif.c b/pif.c
index 8ade587..d028819 100644
--- a/pif.c
+++ b/pif.c
@@ -114,7 +114,7 @@ int pif_listen(const struct ctx *c, uint8_t proto, uint8_t pif,
 		union sockaddr_inany sa;
 
 		pif_sockaddr(c, &sa, pif, addr, port);
-		ref.fd = sock_l4(c, type, &sa, ifname);
+		ref.fd = sock_l4(c, type, &sa, ifname, false);
 	}
 
 	if (ref.fd < 0)
diff --git a/udp_flow.c b/udp_flow.c
index 865702c..a9fab7a 100644
--- a/udp_flow.c
+++ b/udp_flow.c
@@ -74,10 +74,26 @@ static int udp_flow_sock(const struct ctx *c,
 {
 	const struct flowside *side = &uflow->f.side[sidei];
 	uint8_t pif = uflow->f.pif[sidei];
+	/* A guest-initiated flow keeps the guest's source port on the host
+	 * side.  Bind it exclusively: without SO_REUSEADDR the bind fails if
+	 * any other socket in this network namespace holds the port, such as
+	 * another passt instance's flow towards the same destination (typically
+	 * the same DNS resolver).  Two sockets with an identical 4-tuple would
+	 * steal each other's replies.  In that case let the kernel pick a free
+	 * port instead, and udp_flow_new() records it.  Listening sockets and
+	 * the flows they initiate keep SO_REUSEADDR, as they must share ports.
+	 */
+	bool exclusive = sidei == TGTSIDE && uflow->f.pif[INISIDE] == PIF_TAP;
 	int rc;
 	int s;
 
-	s = flowside_sock_l4(c, EPOLL_TYPE_UDP, pif, side);
+	s = flowside_sock_l4(c, EPOLL_TYPE_UDP, pif, side, exclusive);
+	if (s == -EADDRINUSE && exclusive && side->oport) {
+		flow_dbg(uflow, "Host port %hu in use, using an ephemeral port",
+			 side->oport);
+		uflow->f.side[sidei].oport = 0;
+		s = flowside_sock_l4(c, EPOLL_TYPE_UDP, pif, side, exclusive);
+	}
 	if (s < 0) {
 		flow_perror_ratelimit(uflow, now,
 				      "Couldn't open flow specific socket");
@@ -162,13 +178,15 @@ static flow_sidx_t udp_flow_new(const struct ctx *c, union flow *flow,
 				goto cancel;
 	}
 
-	if (uflow->s[TGTSIDE] >= 0 && inany_is_unspecified(&tgt->oaddr)) {
+	if (uflow->s[TGTSIDE] >= 0 &&
+	    (inany_is_unspecified(&tgt->oaddr) || !tgt->oport)) {
 		/* When we target a socket, we connect() it, but might not
 		 * always bind(), leaving the kernel to pick our address.  In
 		 * that case connect() will implicitly bind() the socket, but we
 		 * need to determine its local address so that we can match
 		 * reply packets back to the correct flow.  Update the flow with
-		 * the information from getsockname() */
+		 * the information from getsockname().  The same applies to the
+		 * port if udp_flow_sock() had to let the kernel pick one. */
 		union sockaddr_inany sa;
 		socklen_t sl = sizeof(sa);
 		in_port_t port;
@@ -179,7 +197,9 @@ static flow_sidx_t udp_flow_new(const struct ctx *c, union flow *flow,
 			flow_perror(uflow, "Unable to determine local address");
 			goto cancel;
 		}
-		if (port != tgt->oport) {
+		if (!tgt->oport) {
+			uflow->f.side[TGTSIDE].oport = port;
+		} else if (port != tgt->oport) {
 			flow_err_ratelimit(uflow, now, "Unexpected local port");
 			goto cancel;
 		}
diff --git a/util.c b/util.c
index ff1f19f..053dd80 100644
--- a/util.c
+++ b/util.c
@@ -47,12 +47,14 @@ uint8_t eth_pad[ETH_ZLEN] = { 0 };
  * @sa:		Socket address to bind to
  * @ifname:	Interface for binding, NULL for any
  * @v6only:	If >= 0, set IPV6_V6ONLY socket option to this value
+ * @exclusive:	Don't set SO_REUSEADDR, so that binding a port already bound
+ *		by any other socket fails with EADDRINUSE
  *
  * Return: newly created socket, negative error code on failure
  */
 static int sock_l4_(const struct ctx *c, enum epoll_type type,
 		    const union sockaddr_inany *sa, const char *ifname,
-		    int v6only)
+		    int v6only, bool exclusive)
 {
 	sa_family_t af = sa->sa_family;
 	bool freebind = false;
@@ -104,7 +106,8 @@ static int sock_l4_(const struct ctx *c, enum epoll_type type,
 		}
 	}
 
-	if (setsockopt(fd, SOL_SOCKET, SO_REUSEADDR, &y, sizeof(y)))
+	if (!exclusive &&
+	    setsockopt(fd, SOL_SOCKET, SO_REUSEADDR, &y, sizeof(y)))
 		debug("Failed to set SO_REUSEADDR on socket %i", fd);
 
 	if (proto == IPPROTO_UDP) {
@@ -173,11 +176,13 @@ static int sock_l4_(const struct ctx *c, enum epoll_type type,
  * @type:	epoll type
  * @sa:		Socket address to bind to
  * @ifname:	Interface for binding, NULL for any
+ * @exclusive:	Don't set SO_REUSEADDR (see sock_l4_())
  *
  * Return: newly created socket, negative error code on failure
  */
 int sock_l4(const struct ctx *c, enum epoll_type type,
-	    const union sockaddr_inany *sa, const char *ifname)
+	    const union sockaddr_inany *sa, const char *ifname,
+	    bool exclusive)
 {
 	int v6only = -1;
 
@@ -188,7 +193,7 @@ int sock_l4(const struct ctx *c, enum epoll_type type,
 	    IN6_IS_ADDR_UNSPECIFIED(&sa->sa6.sin6_addr))
 		v6only = 1;
 
-	return sock_l4_(c, type, sa, ifname, v6only);
+	return sock_l4_(c, type, sa, ifname, v6only, exclusive);
 }
 
 /**
@@ -215,7 +220,7 @@ int sock_l4_dualstack_any(const struct ctx *c, enum epoll_type type,
 	 * default, but sysctl net.ipv6.bindv6only can change that, so set the
 	 * sockopt explicitly.
 	 */
-	return sock_l4_(c, type, &sa, ifname, 0);
+	return sock_l4_(c, type, &sa, ifname, 0, false);
 }
 
 /**
diff --git a/util.h b/util.h
index ec2d9ac..0f187ad 100644
--- a/util.h
+++ b/util.h
@@ -147,7 +147,8 @@ struct ctx;
 union sockaddr_inany;
 
 int sock_l4(const struct ctx *c, enum epoll_type type,
-	    const union sockaddr_inany *sa, const char *ifname);
+	    const union sockaddr_inany *sa, const char *ifname,
+	    bool exclusive);
 int sock_l4_dualstack_any(const struct ctx *c, enum epoll_type type,
 			  in_port_t port, const char *ifname);
 int sock_unix(char *sock_path);
-- 
2.43.0


                 reply	other threads:[~2026-11-15 22:38 UTC|newest]

Thread overview: [no followups] expand[flat|nested]  mbox.gz  Atom feed

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20261115223823.1401886-1-sbrivio@redhat.com \
    --to=sbrivio@redhat.com \
    --cc=david@gibson.dropbear.id.au \
    --cc=info@gerwim.nl \
    --cc=passt-dev@passt.top \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
Code repositories for project(s) associated with this public inbox

	https://passt.top/passt

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for IMAP folder(s).