net: phonet: Fix Use-After-Free in pep_recvmsg
Subsystems (3)
PATCH 1/1 net: phonet: Fix Use-After-Free in pep_recvmsg
Because pep_recvmsg() fetches the skb from pn->ctrlreq_queue without holding the lock_sock and then frees it,
a race can occur with pep_ioctl().
A use-after-free for a skb occurs with the following flow.
``` pep_recvmsg() -> skb_dequeue() -> skb_free_datagram() pep_ioctl() -> skb_peek() ```
Fix this by adjusting the scope of lock_sock in pep_recvmsg().
--- net/phonet/pep.c | 17 +++++++++++++---- 1 file changed, 13 insertions(+), 4 deletions(-)
net/phonet/pep.c +13 -4
@@ -1250,12 +1250,17 @@ static int pep_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,if (unlikely (1 << sk->sk_state & (TCPF_LISTEN |TCPF_CLOSE )))return -ENOTCONN ; +lock_sock (sk); +if ((flags &MSG_OOB ) ||sock_flag (sk,SOCK_URGINLINE )) {/* Dequeue and acknowledge control request */ struct pep_sock *pn =pep_sk (sk); -if (flags &MSG_PEEK ) +if (flags &MSG_PEEK ) { +release_sock (sk);return -EOPNOTSUPP ; + } + skb =skb_dequeue (&pn->ctrlreq_queue);if (skb) { pep_ctrlreq_error(sk, skb, PN_PIPE_NO_ERROR,
@@ -1263,12 +1268,14 @@ static int pep_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, msg->msg_flags |=MSG_OOB ; goto copy; } -if (flags &MSG_OOB ) + +if (flags &MSG_OOB ) { +release_sock (sk);return -EINVAL ; + } } skb =skb_recv_datagram (sk, flags, &err); -lock_sock (sk);if (skb ==NULL ) {if (err == -ENOTCONN && sk->sk_state ==TCP_CLOSE_WAIT ) err = -ECONNRESET ;
@@ -1278,7 +1285,7 @@ static int pep_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,if (sk->sk_state ==TCP_ESTABLISHED )pipe_grant_credits (sk,GFP_KERNEL ); -release_sock (sk); + copy: msg->msg_flags |=MSG_EOR ;if (skb->len > len)
@@ -1291,6 +1298,8 @@ static int pep_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, err = (flags &MSG_TRUNC ) ? skb->len : len;skb_free_datagram (sk, skb); + +release_sock (sk);return err; }
Signature · 1 line
2.25.1
Hi,
Le 4 décembre 2023 08:59:52 GMT+02:00, Hyunwoo Kim <[email protected]> a écrit :
Quoted 4 lines
Because
pep_recvmsg()fetches the skb frompn->ctrlreq_queuewithout holding thelock_sockand then frees it,
a race can occur withpep_ioctl().
A use-after-free for a skb occurs with the following flow.
Isn't this the same issue that was reported by Huawei rootlab and for which I already provided a pair of patches to the security list two months ago?
TBH, I much prefer the approach in the other patch set, which takes the hit on the ioctl() side rather than the recvmsg()'s.
Unfortunately, I have no visibility on what happened or didn't happen after that, since the security list is private.
Quoted 21 lines
``` pep_recvmsg() -> skb_dequeue() -> skb_free_datagram() pep_ioctl() -> skb_peek() ```Fix this by adjusting the scope of
lock_sockinpep_recvmsg().Signed-off-by Hyunwoo Kim <[email protected]>---
net/phonet/pep.c| 17 +++++++++++++----1 file changed, 13 insertions(+), 4 deletions(-)net/phonet/pep.c +2
-0@@ -1250,12 +1250,17 @@ static int pep_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,if (unlikely (1 << sk->sk_state & (TCPF_LISTEN |TCPF_CLOSE )))return -ENOTCONN ; +lock_sock (sk); +
net/phonet/pep.c static int pep_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, · line 1263
Quoted 10 lines
net/phonet/pep.c +4
-1if ((flags &MSG_OOB ) ||sock_flag (sk,SOCK_URGINLINE )) {/* Dequeue and acknowledge control request */ struct pep_sock *pn =pep_sk (sk); -if (flags &MSG_PEEK ) +if (flags &MSG_PEEK ) { +release_sock (sk);return -EOPNOTSUPP ; + } +
Also this change is not really accounted for.
Quoted 38 lines
net/phonet/pep.c +5
-3skb =skb_dequeue (&pn->ctrlreq_queue);if (skb) { pep_ctrlreq_error(sk, skb, PN_PIPE_NO_ERROR,@@ -1263,12 +1268,14 @@ static int pep_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, msg->msg_flags |=MSG_OOB ; goto copy; } -if (flags &MSG_OOB ) + +if (flags &MSG_OOB ) { +release_sock (sk);return -EINVAL ; + } } skb =skb_recv_datagram (sk, flags, &err); -lock_sock (sk);if (skb ==NULL ) {if (err == -ENOTCONN && sk->sk_state ==TCP_CLOSE_WAIT ) err = -ECONNRESET ;@@ -1278,7 +1285,7 @@ static int pep_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,if (sk->sk_state ==TCP_ESTABLISHED )pipe_grant_credits (sk,GFP_KERNEL ); -release_sock (sk); +copy:
msg->msg_flags |= MSG_EOR; if (skb->len > len)net/phonet/pep.c +2
-0@@ -1291,6 +1298,8 @@ static int pep_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, err = (flags &MSG_TRUNC ) ? skb->len : len;skb_free_datagram (sk, skb); + +release_sock (sk);return err; }
Hi,
Quoted 9 lines · Rémi Denis-Courmont
Hi,
Le 4 décembre 2023 08:59:52 GMT+02:00, Hyunwoo Kim <[email protected]> a écrit :
Quoted 4 lines
Because
pep_recvmsg()fetches the skb frompn->ctrlreq_queuewithout holding thelock_sockand then frees it,
a race can occur withpep_ioctl().
A use-after-free for a skb occurs with the following flow.Isn't this the same issue that was reported by Huawei rootlab and for which I already provided a pair of patches to the security list two months ago?
Is the issue reported to the security mailing list two months ago the same as this pn->ctrlreq_queue race?
Quoted 2 lines
TBH, I much prefer the approach in the other patch set, which takes the hit on the
ioctl()side rather than the recvmsg()'s.
That's probably a patch to add sk->sk_receive_queue.lock to pep_ioctl(), is that correct?
Quoted 2 lines
Unfortunately, I have no visibility on what happened or didn't happen after that, since the security list is private.
Perhaps this issue hasn't gotten much attention.
Regards,
Hyunwoo Kim
Quoted 73 lines
Quoted 22 lines
``` pep_recvmsg() -> skb_dequeue() -> skb_free_datagram() pep_ioctl() -> skb_peek() ```Fix this by adjusting the scope of
lock_sockinpep_recvmsg().Signed-off-by Hyunwoo Kim <[email protected]>---
net/phonet/pep.c| 17 +++++++++++++----1 file changed, 13 insertions(+), 4 deletions(-)net/phonet/pep.c +2
-0@@ -1250,12 +1250,17 @@ static int pep_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,if (unlikely (1 << sk->sk_state & (TCPF_LISTEN |TCPF_CLOSE )))return -ENOTCONN ; +lock_sock (sk); +net/phonet/pep.c static int pep_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, · line 1263
Quoted 10 lines
net/phonet/pep.c +4
-1if ((flags &MSG_OOB ) ||sock_flag (sk,SOCK_URGINLINE )) {/* Dequeue and acknowledge control request */ struct pep_sock *pn =pep_sk (sk); -if (flags &MSG_PEEK ) +if (flags &MSG_PEEK ) { +release_sock (sk);return -EOPNOTSUPP ; + } +Also this change is not really accounted for.
Quoted 38 lines
net/phonet/pep.c +5
-3skb =skb_dequeue (&pn->ctrlreq_queue);if (skb) { pep_ctrlreq_error(sk, skb, PN_PIPE_NO_ERROR,@@ -1263,12 +1268,14 @@ static int pep_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, msg->msg_flags |=MSG_OOB ; goto copy; } -if (flags &MSG_OOB ) + +if (flags &MSG_OOB ) { +release_sock (sk);return -EINVAL ; + } } skb =skb_recv_datagram (sk, flags, &err); -lock_sock (sk);if (skb ==NULL ) {if (err == -ENOTCONN && sk->sk_state ==TCP_CLOSE_WAIT ) err = -ECONNRESET ;@@ -1278,7 +1285,7 @@ static int pep_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,if (sk->sk_state ==TCP_ESTABLISHED )pipe_grant_credits (sk,GFP_KERNEL ); -release_sock (sk); +copy:
msg->msg_flags |= MSG_EOR; if (skb->len > len)net/phonet/pep.c +2
-0@@ -1291,6 +1298,8 @@ static int pep_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, err = (flags &MSG_TRUNC ) ? skb->len : len;skb_free_datagram (sk, skb); + +release_sock (sk);return err; }
Quoted 6 lines · Le keskiviikkona 6. joulukuuta 2023, 6.25.19 EET Hyunwoo Kim a écrit :
Hi,
Quoted 3 lines · Rémi Denis-Courmont
Hi,
Le 4 décembre 2023 08:59:52 GMT+02:00, Hyunwoo Kim <[email protected]> a
écrit :
Quoted 10 lines
Quoted 8 lines
Quoted 4 lines
Because
pep_recvmsg()fetches the skb frompn->ctrlreq_queuewithout holding thelock_sockand then frees it,
a race can occur withpep_ioctl().
A use-after-free for a skb occurs with the following flow.Isn't this the same issue that was reported by Huawei rootlab and for which I already provided a pair of patches to the security list two months ago?
Is the issue reported to the security mailing list two months ago the same as this
pn->ctrlreq_queuerace?
No, it was another similar problem but the fixes did cover both, I think?
Quoted 4 lines
Quoted 2 lines
TBH, I much prefer the approach in the other patch set, which takes the hit on the
ioctl()side rather than the recvmsg()'s.That's probably a patch to add
sk->sk_receive_queue.locktopep_ioctl(), is that correct?
More or less
Quoted 3 lines
Quoted 2 lines
Unfortunately, I have no visibility on what happened or didn't happen after that, since the security list is private.
Perhaps this issue hasn't gotten much attention.
Quite possible, but now I'm between a rock and a hard place, because I don't know what's (not) going in the security mailing list. In my understanding, it was not really OK to bring the issue or post the patches on netdev :shrug:
Signature · 2 lines
雷米‧德尼-库尔蒙 http://www.remlab.net/