[PATCH] net: Avoid address overwrite in kernel_connect

BPF programs that run on connect can rewrite the connect address. For
the connect system call this isn't a problem, because a copy of the address
is made when it is moved into kernel space. However, kernel_connect
simply passes through the address it is given, so the caller may observe
its address value unexpectedly change.

A practical example where this is problematic is where NFS is combined
with a system such as Cilium which implements BPF-based load balancing.
A common pattern in software-defined storage systems is to have an NFS
mount that connects to a persistent virtual IP which in turn maps to an
ephemeral server IP. This is usually done to achieve high availability:
if your server goes down you can quickly spin up a replacement and remap
the virtual IP to that endpoint. With BPF-based load balancing, mounts
will forget the virtual IP address when the address rewrite occurs
because a pointer to the only copy of that address is passed down the
stack. Server failover then breaks, because clients have forgotten the
virtual IP address. Reconnects fail and mounts remain broken. This patch
was tested by setting up a scenario like this and ensuring that NFS
reconnects worked after applying the patch.

Signed-off-by: Jordan Rife <redacted>
---
 net/socket.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/net/socket.c b/net/socket.c
index 2b0e54b2405c8..f49edb9b49185 100644
--- a/net/socket.c
+++ b/net/socket.c

@@ -3519,7 +3519,11 @@ EXPORT_SYMBOL(kernel_accept);
 int kernel_connect(struct socket *sock, struct sockaddr *addr, int addrlen,
 		   int flags)
 {
-	return sock->ops->connect(sock, addr, addrlen, flags);
+	struct sockaddr_storage address;
+
+	memcpy(&address, addr, addrlen);
+
+	return sock->ops->connect(sock, (struct sockaddr *)&address, addrlen, flags);
 }
 EXPORT_SYMBOL(kernel_connect);
 

-- 
2.42.0.rc1.204.g551eb34607-goog

From: Jordan Rife <redacted>
Date: Mon, 21 Aug 2023 05:00:06 -0500

Could you rebase on net-next.git ?  I think this patch conflicts with
1ded5e5a5931 ("net: annotate data-races around sock->ops").

 }
 EXPORT_SYMBOL(kernel_connect);
 
-- 
2.42.0.rc1.204.g551eb34607-goog

BPF programs that run on connect can rewrite the connect address. For
the connect system call this isn't a problem, because a copy of the address
is made when it is moved into kernel space. However, kernel_connect
simply passes through the address it is given, so the caller may observe
its address value unexpectedly change.

A practical example where this is problematic is where NFS is combined
with a system such as Cilium which implements BPF-based load balancing.
A common pattern in software-defined storage systems is to have an NFS
mount that connects to a persistent virtual IP which in turn maps to an
ephemeral server IP. This is usually done to achieve high availability:
if your server goes down you can quickly spin up a replacement and remap
the virtual IP to that endpoint. With BPF-based load balancing, mounts
will forget the virtual IP address when the address rewrite occurs
because a pointer to the only copy of that address is passed down the
stack. Server failover then breaks, because clients have forgotten the
virtual IP address. Reconnects fail and mounts remain broken. This patch
was tested by setting up a scenario like this and ensuring that NFS
reconnects worked after applying the patch.

Signed-off-by: Jordan Rife <redacted>
---
V1 -> V2: Rebased on net-next

 net/socket.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/net/socket.c b/net/socket.c
index fdb5233bf560c..90c07148835e6 100644
--- a/net/socket.c
+++ b/net/socket.c

@@ -3567,7 +3567,11 @@ EXPORT_SYMBOL(kernel_accept);
 int kernel_connect(struct socket *sock, struct sockaddr *addr, int addrlen,
 		   int flags)
 {
-	return READ_ONCE(sock->ops)->connect(sock, addr, addrlen, flags);
+	struct sockaddr_storage address;
+
+	memcpy(&address, addr, addrlen);
+
+	return READ_ONCE(sock->ops)->connect(sock, (struct sockaddr *)&address, addrlen, flags);
 }
 EXPORT_SYMBOL(kernel_connect);
 

-- 
2.42.0.rc1.204.g551eb34607-goog

BPF programs that run on connect can rewrite the connect address. For
the connect system call this isn't a problem, because a copy of the address
is made when it is moved into kernel space. However, kernel_connect
simply passes through the address it is given, so the caller may observe
its address value unexpectedly change.

A practical example where this is problematic is where NFS is combined
with a system such as Cilium which implements BPF-based load balancing.
A common pattern in software-defined storage systems is to have an NFS
mount that connects to a persistent virtual IP which in turn maps to an
ephemeral server IP. This is usually done to achieve high availability:
if your server goes down you can quickly spin up a replacement and remap
the virtual IP to that endpoint. With BPF-based load balancing, mounts
will forget the virtual IP address when the address rewrite occurs
because a pointer to the only copy of that address is passed down the
stack. Server failover then breaks, because clients have forgotten the
virtual IP address. Reconnects fail and mounts remain broken. This patch
was tested by setting up a scenario like this and ensuring that NFS
reconnects worked after applying the patch.

Signed-off-by: Jordan Rife <redacted>
---
V2 -> V3: Broke up long line
V1 -> V2: Rebased on net-next

 net/socket.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/net/socket.c b/net/socket.c
index fdb5233bf560c..848116d06b511 100644
--- a/net/socket.c
+++ b/net/socket.c

@@ -3567,7 +3567,12 @@ EXPORT_SYMBOL(kernel_accept);
 int kernel_connect(struct socket *sock, struct sockaddr *addr, int addrlen,
 		   int flags)
 {
-	return READ_ONCE(sock->ops)->connect(sock, addr, addrlen, flags);
+	struct sockaddr_storage address;
+
+	memcpy(&address, addr, addrlen);
+
+	return READ_ONCE(sock->ops)->connect(sock, (struct sockaddr *)&address,
+					     addrlen, flags);
 }
 EXPORT_SYMBOL(kernel_connect);
 

-- 
2.42.0.rc1.204.g551eb34607-goog

Hello:

This patch was applied to netdev/net-next.git (main)
by David S. Miller [off-list ref]:

On Mon, 21 Aug 2023 16:45:23 -0500 you wrote:

BPF programs that run on connect can rewrite the connect address. For
the connect system call this isn't a problem, because a copy of the address
is made when it is moved into kernel space. However, kernel_connect
simply passes through the address it is given, so the caller may observe
its address value unexpectedly change.

A practical example where this is problematic is where NFS is combined
with a system such as Cilium which implements BPF-based load balancing.
A common pattern in software-defined storage systems is to have an NFS
mount that connects to a persistent virtual IP which in turn maps to an
ephemeral server IP. This is usually done to achieve high availability:
if your server goes down you can quickly spin up a replacement and remap
the virtual IP to that endpoint. With BPF-based load balancing, mounts
will forget the virtual IP address when the address rewrite occurs
because a pointer to the only copy of that address is passed down the
stack. Server failover then breaks, because clients have forgotten the
virtual IP address. Reconnects fail and mounts remain broken. This patch
was tested by setting up a scenario like this and ensuring that NFS
reconnects worked after applying the patch.

[...]

Here is the summary with links:
  - [net-next,v3] net: Avoid address overwrite in kernel_connect
    https://git.kernel.org/netdev/net-next/c/0bdf399342c5

You are awesome, thank you!
-- 
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html