[PATCH v2] net/smc: fix TCP fallback socket release

clcsock can be released while kernel_accept() references it in TCP
listen worker. Also, clcsock needs to wake up before released if TCP
fallback is used and the clcsock is blocked by accept. Add a lock to
safely release clcsock and call kernel_sock_shutdown() to wake up
clcsock from accept in smc_release().

Reported-by: syzbot+0bf2e01269f1274b4b03@syzkaller.appspotmail.com
Reported-by: syzbot+e3132895630f957306bc@syzkaller.appspotmail.com
Signed-off-by: Myungho Jung <redacted>
---
 net/smc/af_smc.c | 14 ++++++++++++--
 net/smc/smc.h    |  2 ++
 2 files changed, 14 insertions(+), 2 deletions(-)

diff --git a/net/smc/af_smc.c b/net/smc/af_smc.c
index 5fbaf1901571..5d06fb1bbccf 100644
--- a/net/smc/af_smc.c
+++ b/net/smc/af_smc.c

@@ -147,8 +147,14 @@ static int smc_release(struct socket *sock)
 		sk->sk_shutdown |= SHUTDOWN_MASK;
 	}
 	if (smc->clcsock) {
+		if (smc->use_fallback && sk->sk_state == SMC_LISTEN) {
+			/* wake up clcsock accept */
+			rc = kernel_sock_shutdown(smc->clcsock, SHUT_RDWR);
+		}
+		mutex_lock(&smc->clcsock_release_lock);
 		sock_release(smc->clcsock);
 		smc->clcsock = NULL;
+		mutex_unlock(&smc->clcsock_release_lock);
 	}
 	if (smc->use_fallback) {
 		if (sk->sk_state != SMC_LISTEN && sk->sk_state != SMC_INIT)

@@ -205,6 +211,7 @@ static struct sock *smc_sock_alloc(struct net *net, struct socket *sock,
 	spin_lock_init(&smc->conn.send_lock);
 	sk->sk_prot->hash(sk);
 	sk_refcnt_debug_inc(sk);
+	mutex_init(&smc->clcsock_release_lock);
 
 	return sk;
 }

@@ -821,7 +828,7 @@ static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
 	struct socket *new_clcsock = NULL;
 	struct sock *lsk = &lsmc->sk;
 	struct sock *new_sk;
-	int rc;
+	int rc = 0;
 
 	release_sock(lsk);
 	new_sk = smc_sock_alloc(sock_net(lsk), NULL, lsk->sk_protocol);

@@ -834,7 +841,10 @@ static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
 	}
 	*new_smc = smc_sk(new_sk);
 
-	rc = kernel_accept(lsmc->clcsock, &new_clcsock, 0);
+	mutex_lock(&lsmc->clcsock_release_lock);
+	if (lsmc->clcsock)
+		rc = kernel_accept(lsmc->clcsock, &new_clcsock, 0);
+	mutex_unlock(&lsmc->clcsock_release_lock);
 	lock_sock(lsk);
 	if  (rc < 0)
 		lsk->sk_err = -rc;

diff --git a/net/smc/smc.h b/net/smc/smc.h
index 08786ace6010..9a2795cf5d30 100644
--- a/net/smc/smc.h
+++ b/net/smc/smc.h

@@ -219,6 +219,8 @@ struct smc_sock {				/* smc sock container */
 						 * started, waiting for unsent
 						 * data to be sent
 						 */
+	struct mutex            clcsock_release_lock;
+						/* protects clcsock */
 };
 
 static inline struct smc_sock *smc_sk(const struct sock *sk)

-- 
2.17.1

On 12/17/2018 06:21 AM, Myungho Jung wrote:

clcsock can be released while kernel_accept() references it in TCP
listen worker. Also, clcsock needs to wake up before released if TCP
fallback is used and the clcsock is blocked by accept. Add a lock to
safely release clcsock and call kernel_sock_shutdown() to wake up
clcsock from accept in smc_release().

Thanks for your effort to solve this problem. I have some minor
improvement proposals:

This part is not needed, since an SMC socket in state SMC_LISTEN is never
a use_fallback socket.

Without clcsock the good path should not be executed. Thus I suggest
to initialize with something negative like -EINVAL.

I suggest to be more precise: "protects clcsock of a listen socket" 

 };
 
 static inline struct smc_sock *smc_sk(const struct sock *sk)

On Mon, Dec 17, 2018 at 03:58:58PM +0100, Ursula Braun wrote:

Hi Ursula,

Thank you for your suggestion. I have a question on your comment.

On 12/17/2018 06:21 AM, Myungho Jung wrote:

quoted

clcsock can be released while kernel_accept() references it in TCP
listen worker. Also, clcsock needs to wake up before released if TCP
fallback is used and the clcsock is blocked by accept. Add a lock to
safely release clcsock and call kernel_sock_shutdown() to wake up
clcsock from accept in smc_release().

Thanks for your effort to solve this problem. I have some minor
improvement proposals:

quoted

Reported-by: syzbot+0bf2e01269f1274b4b03@syzkaller.appspotmail.com
Reported-by: syzbot+e3132895630f957306bc@syzkaller.appspotmail.com
Signed-off-by: Myungho Jung <redacted>
---
 net/smc/af_smc.c | 14 ++++++++++++--
 net/smc/smc.h    |  2 ++
 2 files changed, 14 insertions(+), 2 deletions(-)

diff --git a/net/smc/af_smc.c b/net/smc/af_smc.c
index 5fbaf1901571..5d06fb1bbccf 100644
--- a/net/smc/af_smc.c
+++ b/net/smc/af_smc.c

@@ -147,8 +147,14 @@ static int smc_release(struct socket *sock)
 		sk->sk_shutdown |= SHUTDOWN_MASK;
 	}
 	if (smc->clcsock) {
+		if (smc->use_fallback && sk->sk_state == SMC_LISTEN) {
+			/* wake up clcsock accept */
+			rc = kernel_sock_shutdown(smc->clcsock, SHUT_RDWR);
+		}

This part is not needed, since an SMC socket in state SMC_LISTEN is never
a use_fallback socket.

In smc_sendmsg(), set use_fallback to true if SMC socket is SMC_INIT
state and the message has MSG_FASTOPEN flag. After this, smc_listen()
would trigger smc_tcp_listen_work(). Is this not an expected scenario?
Then, what is the reason for not skipping smc_sendmsg() in SMC_INIT
state?

quoted

+		mutex_lock(&smc->clcsock_release_lock);
 		sock_release(smc->clcsock);
 		smc->clcsock = NULL;
+		mutex_unlock(&smc->clcsock_release_lock);
 	}
 	if (smc->use_fallback) {
 		if (sk->sk_state != SMC_LISTEN && sk->sk_state != SMC_INIT)

@@ -205,6 +211,7 @@ static struct sock *smc_sock_alloc(struct net *net, struct socket *sock,
 	spin_lock_init(&smc->conn.send_lock);
 	sk->sk_prot->hash(sk);
 	sk_refcnt_debug_inc(sk);
+	mutex_init(&smc->clcsock_release_lock);
 
 	return sk;
 }

@@ -821,7 +828,7 @@ static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
 	struct socket *new_clcsock = NULL;
 	struct sock *lsk = &lsmc->sk;
 	struct sock *new_sk;
-	int rc;
+	int rc = 0;

Without clcsock the good path should not be executed. Thus I suggest
to initialize with something negative like -EINVAL.

quoted

 
 	release_sock(lsk);
 	new_sk = smc_sock_alloc(sock_net(lsk), NULL, lsk->sk_protocol);

@@ -834,7 +841,10 @@ static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
 	}
 	*new_smc = smc_sk(new_sk);
 
-	rc = kernel_accept(lsmc->clcsock, &new_clcsock, 0);
+	mutex_lock(&lsmc->clcsock_release_lock);
+	if (lsmc->clcsock)
+		rc = kernel_accept(lsmc->clcsock, &new_clcsock, 0);
+	mutex_unlock(&lsmc->clcsock_release_lock);
 	lock_sock(lsk);
 	if  (rc < 0)
 		lsk->sk_err = -rc;

diff --git a/net/smc/smc.h b/net/smc/smc.h
index 08786ace6010..9a2795cf5d30 100644
--- a/net/smc/smc.h
+++ b/net/smc/smc.h

@@ -219,6 +219,8 @@ struct smc_sock {				/* smc sock container */
 						 * started, waiting for unsent
 						 * data to be sent
 						 */
+	struct mutex            clcsock_release_lock;
+						/* protects clcsock */

I suggest to be more precise: "protects clcsock of a listen socket" 

quoted

 };
 
 static inline struct smc_sock *smc_sk(const struct sock *sk)

On 12/18/2018 08:03 AM, Myungho Jung wrote:

On Mon, Dec 17, 2018 at 03:58:58PM +0100, Ursula Braun wrote:

quoted

Hi Ursula,

Thank you for your suggestion. I have a question on your comment.

quoted

On 12/17/2018 06:21 AM, Myungho Jung wrote:

quoted

clcsock can be released while kernel_accept() references it in TCP
listen worker. Also, clcsock needs to wake up before released if TCP
fallback is used and the clcsock is blocked by accept. Add a lock to
safely release clcsock and call kernel_sock_shutdown() to wake up
clcsock from accept in smc_release().

Thanks for your effort to solve this problem. I have some minor
improvement proposals:

quoted

Reported-by: syzbot+0bf2e01269f1274b4b03@syzkaller.appspotmail.com
Reported-by: syzbot+e3132895630f957306bc@syzkaller.appspotmail.com
Signed-off-by: Myungho Jung <redacted>
---
 net/smc/af_smc.c | 14 ++++++++++++--
 net/smc/smc.h    |  2 ++
 2 files changed, 14 insertions(+), 2 deletions(-)

diff --git a/net/smc/af_smc.c b/net/smc/af_smc.c
index 5fbaf1901571..5d06fb1bbccf 100644
--- a/net/smc/af_smc.c
+++ b/net/smc/af_smc.c

@@ -147,8 +147,14 @@ static int smc_release(struct socket *sock)
 		sk->sk_shutdown |= SHUTDOWN_MASK;
 	}
 	if (smc->clcsock) {
+		if (smc->use_fallback && sk->sk_state == SMC_LISTEN) {
+			/* wake up clcsock accept */
+			rc = kernel_sock_shutdown(smc->clcsock, SHUT_RDWR);
+		}

This part is not needed, since an SMC socket in state SMC_LISTEN is never
a use_fallback socket.

In smc_sendmsg(), set use_fallback to true if SMC socket is SMC_INIT
state and the message has MSG_FASTOPEN flag. After this, smc_listen()
would trigger smc_tcp_listen_work(). Is this not an expected scenario?
Then, what is the reason for not skipping smc_sendmsg() in SMC_INIT
state?

You are right, I have not had the FASTOPEN case in mind, sorry. If we want
to allow fallback in case of FASTOPEN, we need the kernel_sock_shutdown() here
for proper cleanup. Nice!

quoted

quoted

+		mutex_lock(&smc->clcsock_release_lock);
 		sock_release(smc->clcsock);
 		smc->clcsock = NULL;
+		mutex_unlock(&smc->clcsock_release_lock);
 	}
 	if (smc->use_fallback) {
 		if (sk->sk_state != SMC_LISTEN && sk->sk_state != SMC_INIT)

@@ -205,6 +211,7 @@ static struct sock *smc_sock_alloc(struct net *net, struct socket *sock,
 	spin_lock_init(&smc->conn.send_lock);
 	sk->sk_prot->hash(sk);
 	sk_refcnt_debug_inc(sk);
+	mutex_init(&smc->clcsock_release_lock);
 
 	return sk;
 }

@@ -821,7 +828,7 @@ static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
 	struct socket *new_clcsock = NULL;
 	struct sock *lsk = &lsmc->sk;
 	struct sock *new_sk;
-	int rc;
+	int rc = 0;

Without clcsock the good path should not be executed. Thus I suggest
to initialize with something negative like -EINVAL.

quoted

 
 	release_sock(lsk);
 	new_sk = smc_sock_alloc(sock_net(lsk), NULL, lsk->sk_protocol);

@@ -834,7 +841,10 @@ static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
 	}
 	*new_smc = smc_sk(new_sk);
 
-	rc = kernel_accept(lsmc->clcsock, &new_clcsock, 0);
+	mutex_lock(&lsmc->clcsock_release_lock);
+	if (lsmc->clcsock)
+		rc = kernel_accept(lsmc->clcsock, &new_clcsock, 0);
+	mutex_unlock(&lsmc->clcsock_release_lock);
 	lock_sock(lsk);
 	if  (rc < 0)
 		lsk->sk_err = -rc;

diff --git a/net/smc/smc.h b/net/smc/smc.h
index 08786ace6010..9a2795cf5d30 100644
--- a/net/smc/smc.h
+++ b/net/smc/smc.h

@@ -219,6 +219,8 @@ struct smc_sock {				/* smc sock container */
 						 * started, waiting for unsent
 						 * data to be sent
 						 */
+	struct mutex            clcsock_release_lock;
+						/* protects clcsock */

I suggest to be more precise: "protects clcsock of a listen socket" 

quoted

 };
 
 static inline struct smc_sock *smc_sk(const struct sock *sk)