Skip to contentSkip to conversation

ax25: use after free in ax25_connect

PATCH In review Hangyu Hua 2 participants 8 messages Jan 11, 2022 – Jan 17, 2022

Subsystems (2)
lore
Share & tools
lore

Hangyu Hua

author
#0 lore raw

PATCH 1/1 ax25: use after free in ax25_connect

sk_to_ax25(sk) needs to be called after lock_sock(sk) to avoid UAF caused by a race condition.

---
 net/ax25/af_ax25.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
net/ax25/af_ax25.c +3 -1
index cfca99e295b8..c5d62420a2a8 100644
@@ -1127,7 +1127,7 @@ static int __must_check ax25_connect(struct socket *sock,
 	struct sockaddr *uaddr, int addr_len, int flags)
 {
 	struct sock *sk = sock->sk;
-	ax25_cb *ax25 = sk_to_ax25(sk), *ax25t;
+	ax25_cb *ax25, *ax25t;
 	struct full_sockaddr_ax25 *fsa = (struct full_sockaddr_ax25 *)uaddr;
 	ax25_digi *digi = NULL;
 	int ct = 0, err = 0;
@@ -1155,6 +1155,8 @@ static int __must_check ax25_connect(struct socket *sock,
 
 	lock_sock(sk);
 
+	ax25 = sk_to_ax25(sk);
+
 	/* deal with restarts */
 	if (sock->state == SS_CONNECTING) {
 		switch (sk->sk_state) {
Signature · 1 line
2.25.1
#1 lore raw
Quoted 2 lines · Hangyu Hua

sk_to_ax25(sk) needs to be called after lock_sock(sk) to avoid UAF caused by a race condition.

Can you describe what race condition you have found exactly ?

sk pointer can not change.

Quoted 27 lines
---
  net/ax25/af_ax25.c | 4 +++-
  1 file changed, 3 insertions(+), 1 deletion(-)
net/ax25/af_ax25.c +3 -1
index cfca99e295b8..c5d62420a2a8 100644
@@ -1127,7 +1127,7 @@ static int __must_check ax25_connect(struct socket *sock,
  	struct sockaddr *uaddr, int addr_len, int flags)
  {
  	struct sock *sk = sock->sk;
-	ax25_cb *ax25 = sk_to_ax25(sk), *ax25t;
+	ax25_cb *ax25, *ax25t;
  	struct full_sockaddr_ax25 *fsa = (struct full_sockaddr_ax25 *)uaddr;
  	ax25_digi *digi = NULL;
  	int ct = 0, err = 0;
@@ -1155,6 +1155,8 @@ static int __must_check ax25_connect(struct socket *sock,
 
  	lock_sock(sk);
 
+	ax25 = sk_to_ax25(sk);
+
  	/* deal with restarts */
  	if (sock->state == SS_CONNECTING) {
  		switch (sk->sk_state) {
author
#2 lore raw

I try to use ax25_release to trigger this bug like this:
ax25_release ax25_connect
lock_sock(sk);
-----------------------------sk = sock->sk; -----------------------------ax25 = sk_to_ax25(sk); ax25_destroy_socket(ax25);
release_sock(sk); -----------------------------lock_sock(sk); -----------------------------use ax25 again

But i failed beacause their have large speed difference. And i don't have a physical device to test other function in ax25. Anyway, i still think there will have a function to trigger this race condition like ax25_destroy_timer. Beacause Any ohter functions in ax25_proto_ops like ax25_bind protect ax25_sock by lock_sock(sk).

Thanks.

Quoted 37 lines · Eric Dumazet
Quoted 2 lines · Hangyu Hua

sk_to_ax25(sk) needs to be called after lock_sock(sk) to avoid UAF caused by a race condition.

Can you describe what race condition you have found exactly ?

sk pointer can not change.

Quoted 27 lines

---
  net/ax25/af_ax25.c | 4 +++-

  1 file changed, 3 insertions(+), 1 deletion(-)
net/ax25/af_ax25.c +1 -1
index cfca99e295b8..c5d62420a2a8 100644
@@ -1127,7 +1127,7 @@ static int __must_check ax25_connect(struct 
 socket *sock,
       struct sockaddr *uaddr, int addr_len, int flags)
   {
       struct sock *sk = sock->sk;
-    ax25_cb *ax25 = sk_to_ax25(sk), *ax25t;
+    ax25_cb *ax25, *ax25t;
       struct full_sockaddr_ax25 *fsa = (struct full_sockaddr_ax25 *)uaddr;
       ax25_digi *digi = NULL;

      int ct = 0, err = 0;

net/ax25/af_ax25.c +2 -0
@@ -1155,6 +1155,8 @@ static int __must_check ax25_connect(struct 
 socket *sock,
       lock_sock(sk);
+    ax25 = sk_to_ax25(sk);
+
       /* deal with restarts */
       if (sock->state == SS_CONNECTING) {
           switch (sk->sk_state) {
#3 lore raw
Quoted 10 lines · Hangyu Hua

I try to use ax25_release to trigger this bug like this:
ax25_release                 ax25_connect lock_sock(sk);
-----------------------------sk = sock->sk; -----------------------------ax25 = sk_to_ax25(sk); ax25_destroy_socket(ax25);
release_sock(sk); -----------------------------lock_sock(sk); -----------------------------use ax25 again

Quoted 5 lines

But i failed beacause their have large speed difference. And i don't have a physical device to test other function in ax25. Anyway, i still think there will have a function to trigger this race condition like ax25_destroy_timer. Beacause Any ohter functions in ax25_proto_ops like ax25_bind protect ax25_sock by lock_sock(sk).

For a given sk pointer, sk_to_ax25(sk) is always returning the same value,

regardless of sk lock being held or not.

ax25_sk(sk)->cb  is set only from ax25_create() or ax25_make_new()

ax25_connect can not be called until these operations have completed ?

Quoted 44 lines

Thanks.

Quoted 37 lines · Eric Dumazet
Quoted 2 lines · Hangyu Hua

sk_to_ax25(sk) needs to be called after lock_sock(sk) to avoid UAF caused by a race condition.

Can you describe what race condition you have found exactly ?

sk pointer can not change.

Quoted 27 lines

---
  net/ax25/af_ax25.c | 4 +++-

  1 file changed, 3 insertions(+), 1 deletion(-)
net/ax25/af_ax25.c +1 -1
index cfca99e295b8..c5d62420a2a8 100644
@@ -1127,7 +1127,7 @@ static int __must_check ax25_connect(struct 
 socket *sock,
       struct sockaddr *uaddr, int addr_len, int flags)
   {
       struct sock *sk = sock->sk;
-    ax25_cb *ax25 = sk_to_ax25(sk), *ax25t;
+    ax25_cb *ax25, *ax25t;
       struct full_sockaddr_ax25 *fsa = (struct full_sockaddr_ax25 *)uaddr;
       ax25_digi *digi = NULL;

      int ct = 0, err = 0;

net/ax25/af_ax25.c +2 -0
@@ -1155,6 +1155,8 @@ static int __must_check ax25_connect(struct 
 socket *sock,
       lock_sock(sk);
+    ax25 = sk_to_ax25(sk);
+
       /* deal with restarts */
       if (sock->state == SS_CONNECTING) {
           switch (sk->sk_state) {
author
#4 lore raw

Yes.

And there are two ways to release ax25, ax25_release and time expiry. I tested that ax25_release will not be invoked before ax25_connect is done by closing fd from user space. I think the reason is that __sys_connect use fdget() to protect fd. But i can't test if a function like ax25_std_heartbeat_expiry will release ax25 between sk_to_ax25(sk) and lock_sock(sk).

So i think it's better to protect sk_to_ax25(sk) by a lock. Beacause functions like ax25_release use sk_to_ax25 after a lock.

Quoted 73 lines · Eric Dumazet
Quoted 10 lines · Hangyu Hua

I try to use ax25_release to trigger this bug like this:
ax25_release                 ax25_connect lock_sock(sk);
-----------------------------sk = sock->sk; -----------------------------ax25 = sk_to_ax25(sk); ax25_destroy_socket(ax25);
release_sock(sk); -----------------------------lock_sock(sk); -----------------------------use ax25 again

Quoted 5 lines

But i failed beacause their have large speed difference. And i don't have a physical device to test other function in ax25. Anyway, i still think there will have a function to trigger this race condition like ax25_destroy_timer. Beacause Any ohter functions in ax25_proto_ops like ax25_bind protect ax25_sock by lock_sock(sk).

For a given sk pointer, sk_to_ax25(sk) is always returning the same value,

regardless of sk lock being held or not.

ax25_sk(sk)->cb  is set only from ax25_create() or ax25_make_new()

ax25_connect can not be called until these operations have completed ?

Quoted 44 lines

Thanks.

Quoted 37 lines · Eric Dumazet
Quoted 2 lines · Hangyu Hua

sk_to_ax25(sk) needs to be called after lock_sock(sk) to avoid UAF caused by a race condition.

Can you describe what race condition you have found exactly ?

sk pointer can not change.

Quoted 27 lines

---
  net/ax25/af_ax25.c | 4 +++-

  1 file changed, 3 insertions(+), 1 deletion(-)
net/ax25/af_ax25.c +1 -1
index cfca99e295b8..c5d62420a2a8 100644
@@ -1127,7 +1127,7 @@ static int __must_check ax25_connect(struct 
 socket *sock,
       struct sockaddr *uaddr, int addr_len, int flags)
   {
       struct sock *sk = sock->sk;
-    ax25_cb *ax25 = sk_to_ax25(sk), *ax25t;
+    ax25_cb *ax25, *ax25t;
       struct full_sockaddr_ax25 *fsa = (struct full_sockaddr_ax25 *)uaddr;
       ax25_digi *digi = NULL;

      int ct = 0, err = 0;

net/ax25/af_ax25.c +2 -0
@@ -1155,6 +1155,8 @@ static int __must_check ax25_connect(struct 
 socket *sock,
       lock_sock(sk);
+    ax25 = sk_to_ax25(sk);
+
       /* deal with restarts */
       if (sock->state == SS_CONNECTING) {
           switch (sk->sk_state) {
author
#5 lore raw

Any suggestions for this patch ? Guys.

I think putting sk_to_ax25 after lock_sock(sk) here will avoid any possilbe race conditions like other functions in ax25_proto_ops.

Quoted 81 lines · Hangyu Hua

Yes.

And there are two ways to release ax25, ax25_release and time expiry. I tested that ax25_release will not be invoked before ax25_connect is done by closing fd from user space. I think the reason is that __sys_connect use fdget() to protect fd. But i can't test if a function like ax25_std_heartbeat_expiry will release ax25 between sk_to_ax25(sk) and lock_sock(sk).

So i think it's better to protect sk_to_ax25(sk) by a lock. Beacause functions like ax25_release use sk_to_ax25 after a lock.

Quoted 73 lines · Eric Dumazet
Quoted 10 lines · Hangyu Hua

I try to use ax25_release to trigger this bug like this:
ax25_release                 ax25_connect lock_sock(sk);
-----------------------------sk = sock->sk; -----------------------------ax25 = sk_to_ax25(sk); ax25_destroy_socket(ax25);
release_sock(sk); -----------------------------lock_sock(sk); -----------------------------use ax25 again

Quoted 5 lines

But i failed beacause their have large speed difference. And i don't have a physical device to test other function in ax25. Anyway, i still think there will have a function to trigger this race condition like ax25_destroy_timer. Beacause Any ohter functions in ax25_proto_ops like ax25_bind protect ax25_sock by lock_sock(sk).

For a given sk pointer, sk_to_ax25(sk) is always returning the same value,

regardless of sk lock being held or not.

ax25_sk(sk)->cb  is set only from ax25_create() or ax25_make_new()

ax25_connect can not be called until these operations have completed ?

Quoted 44 lines

Thanks.

Quoted 37 lines · Eric Dumazet
Quoted 2 lines · Hangyu Hua

sk_to_ax25(sk) needs to be called after lock_sock(sk) to avoid UAF caused by a race condition.

Can you describe what race condition you have found exactly ?

sk pointer can not change.

Quoted 27 lines

---
  net/ax25/af_ax25.c | 4 +++-

  1 file changed, 3 insertions(+), 1 deletion(-)
net/ax25/af_ax25.c +1 -1
index cfca99e295b8..c5d62420a2a8 100644
@@ -1127,7 +1127,7 @@ static int __must_check ax25_connect(struct 
 socket *sock,
       struct sockaddr *uaddr, int addr_len, int flags)
   {
       struct sock *sk = sock->sk;
-    ax25_cb *ax25 = sk_to_ax25(sk), *ax25t;
+    ax25_cb *ax25, *ax25t;
       struct full_sockaddr_ax25 *fsa = (struct full_sockaddr_ax25 *)uaddr;
       ax25_digi *digi = NULL;

      int ct = 0, err = 0;

net/ax25/af_ax25.c +2 -0
@@ -1155,6 +1155,8 @@ static int __must_check ax25_connect(struct 
 socket *sock,
       lock_sock(sk);
+    ax25 = sk_to_ax25(sk);
+
       /* deal with restarts */
       if (sock->state == SS_CONNECTING) {
           switch (sk->sk_state) {
#6 lore raw
Quoted 4 lines · Hangyu Hua

Any suggestions for this patch ? Guys.

I think putting sk_to_ax25 after lock_sock(sk) here will avoid any possilbe race conditions like other functions in ax25_proto_ops. CTING) {

As explained, your patch is not needed.

You failed to describe how a race was possible.

Just moving code around wont help.

How about providing a stack trace or some syzbot repro ?

author
#7 lore raw

I get it.

Thanks.

Quoted 15 lines · Eric Dumazet
Quoted 4 lines · Hangyu Hua

Any suggestions for this patch ? Guys.

I think putting sk_to_ax25 after lock_sock(sk) here will avoid any possilbe race conditions like other functions in ax25_proto_ops. CTING) {

As explained, your patch is not needed.

You failed to describe how a race was possible.

Just moving code around wont help.

How about providing a stack trace or some syzbot repro ?