[PATCH net] 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.

Signed-off-by: Hangyu Hua <redacted>
---
 net/ax25/af_ax25.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/net/ax25/af_ax25.c b/net/ax25/af_ax25.c
index cfca99e295b8..c5d62420a2a8 100644
--- a/net/ax25/af_ax25.c
+++ b/net/ax25/af_ax25.c

@@ -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) {

-- 
2.25.1

On 1/10/22 20:20, Hangyu Hua wrote:

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.

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.




On 2022/1/12 上午4:56, Eric Dumazet wrote:

On 1/10/22 20:20, Hangyu Hua wrote:

quoted

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

Signed-off-by: Hangyu Hua <redacted>
---
  net/ax25/af_ax25.c | 4 +++-
  1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/net/ax25/af_ax25.c b/net/ax25/af_ax25.c
index cfca99e295b8..c5d62420a2a8 100644
--- a/net/ax25/af_ax25.c
+++ b/net/ax25/af_ax25.c

@@ -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) {

On 1/11/22 18:13, Hangyu Hua wrote:

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).

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 ?

Thanks.




On 2022/1/12 上午4:56, Eric Dumazet wrote:

quoted

On 1/10/22 20:20, Hangyu Hua wrote:

quoted

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

Signed-off-by: Hangyu Hua <redacted>
---
  net/ax25/af_ax25.c | 4 +++-
  1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/net/ax25/af_ax25.c b/net/ax25/af_ax25.c
index cfca99e295b8..c5d62420a2a8 100644
--- a/net/ax25/af_ax25.c
+++ b/net/ax25/af_ax25.c

@@ -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) {

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.


On 2022/1/12 下午5:59, Eric Dumazet wrote:

On 1/11/22 18:13, Hangyu Hua wrote:

quoted

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).

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

Thanks.




On 2022/1/12 上午4:56, Eric Dumazet wrote:

quoted

On 1/10/22 20:20, Hangyu Hua wrote:

quoted

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

Signed-off-by: Hangyu Hua <redacted>
---
  net/ax25/af_ax25.c | 4 +++-
  1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/net/ax25/af_ax25.c b/net/ax25/af_ax25.c
index cfca99e295b8..c5d62420a2a8 100644
--- a/net/ax25/af_ax25.c
+++ b/net/ax25/af_ax25.c

@@ -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) {

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.

On 2022/1/12 下午7:11, Hangyu Hua wrote:

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.


On 2022/1/12 下午5:59, Eric Dumazet wrote:

quoted

On 1/11/22 18:13, Hangyu Hua wrote:

quoted

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).

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

Thanks.




On 2022/1/12 上午4:56, Eric Dumazet wrote:

quoted

On 1/10/22 20:20, Hangyu Hua wrote:

quoted

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

Signed-off-by: Hangyu Hua <redacted>
---
  net/ax25/af_ax25.c | 4 +++-
  1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/net/ax25/af_ax25.c b/net/ax25/af_ax25.c
index cfca99e295b8..c5d62420a2a8 100644
--- a/net/ax25/af_ax25.c
+++ b/net/ax25/af_ax25.c

@@ -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) {

On 1/13/22 22:54, Hangyu Hua wrote:

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 ?

I get it.

Thanks.

On 2022/1/14 下午11:19, Eric Dumazet wrote:

On 1/13/22 22:54, Hangyu Hua wrote:

quoted

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 ?