tun/tap: set sk_uid from current_fsuid()
Subsystems (4)
cover tun/tap: set sk_uid from current_fsuid()
The original patches fixing CVE-2023-1076 are incorrect in my opinion. This small series fixes them up; see the individual commit messages for explanation.
I have a very elaborate test procedure demonstrating the problem for both tun and tap; it involves libvirt, qemu, and "crash". I can share that procedure if necessary, but it's indeed quite long (I wrote it originally for our QE team).
The patches in this series are supposed to "re-fix" CVE-2023-1076; given that said CVE is classified as Low Impact (CVSSv3=5.5), I'm posting this publicly, and not suggesting any embargo. Red Hat Product Security may assign a new CVE number later.
I've tested the patches on top of v6.5-rc4, with "crash" built at commit c74f375e0ef7.
Laszlo Ersek (2):
net: tun_chr_open(): set sk_uid from current_fsuid()
net: tap_open(): set sk_uid from current_fsuid()
drivers/net/tap.c | 2 +- drivers/net/tun.c | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-)
base-commit: 5d0c230f1de8c7515b6567d9afba1f196fb4e2f4
PATCH 1/2 net: tun_chr_open(): set sk_uid from current_fsuid()
Commit a096ccca6e50 initializes the "sk_uid" field in the protocol socket (struct sock) from the "/dev/net/tun" device node's owner UID. Per original commit 86741ec25462 ("net: core: Add a UID field to struct sock.", 2016-11-04), that's wrong: the idea is to cache the UID of the userspace process that creates the socket. Commit 86741ec25462 mentions socket() and accept(); with "tun", the action that creates the socket is open("/dev/net/tun").
Therefore the device node's owner UID is irrelevant. In most cases, "/dev/net/tun" will be owned by root, so in practice, commit a096ccca6e50 has no observable effect:
- before, "sk_uid" would be zero, due to undefined behavior (CVE-2023-1076),
- after, "sk_uid" would be zero, due to "/dev/net/tun" being owned by root.
What matters is the (fs)UID of the process performing the open(), so cache that in "sk_uid".
Bugzilla: https://bugzilla.redhat.com/show_bug.cgi?id=2173435
--- drivers/net/tun.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/net/tun.c +1 -1
@@ -3469,7 +3469,7 @@ static int tun_chr_open(struct inode *inode, struct file * file) tfile->socket.file = file; tfile->socket.ops = &tun_socket_ops; -sock_init_data_uid (&tfile->socket, &tfile->sk, inode->i_uid); +sock_init_data_uid (&tfile->socket, &tfile->sk,current_fsuid ()); tfile->sk.sk_write_space = tun_sock_write_space; tfile->sk.sk_sndbuf =INT_MAX ;
PATCH 2/2 net: tap_open(): set sk_uid from current_fsuid()
In reply to Laszlo ErsekCommit 66b2c338adce initializes the "sk_uid" field in the protocol socket (struct sock) from the "/dev/tapX" device node's owner UID. Per original commit 86741ec25462 ("net: core: Add a UID field to struct sock.", 2016-11-04), that's wrong: the idea is to cache the UID of the userspace process that creates the socket. Commit 86741ec25462 mentions socket() and accept(); with "tap", the action that creates the socket is open("/dev/tapX").
Therefore the device node's owner UID is irrelevant. In most cases, "/dev/tapX" will be owned by root, so in practice, commit 66b2c338adce has no observable effect:
- before, "sk_uid" would be zero, due to undefined behavior (CVE-2023-1076),
- after, "sk_uid" would be zero, due to "/dev/tapX" being owned by root.
What matters is the (fs)UID of the process performing the open(), so cache that in "sk_uid".
Bugzilla: https://bugzilla.redhat.com/show_bug.cgi?id=2173435
--- drivers/net/tap.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/net/tap.c +1 -1
@@ -534,7 +534,7 @@ static int tap_open(struct inode *inode, struct file *file) q->sock.state =SS_CONNECTED ; q->sock.file = file; q->sock.ops = &tap_socket_ops; -sock_init_data_uid (&q->sock, &q->sk, inode->i_uid); +sock_init_data_uid (&q->sock, &q->sk,current_fsuid ()); q->sk.sk_write_space = tap_sock_write_space; q->sk.sk_destruct = tap_sock_destruct; q->flags =IFF_VNET_HDR |IFF_NO_PI |IFF_TAP ;
Jakub Kicinski
maintainerQuoted 9 lines · On Mon, 31 Jul 2023 18:42:35 +0200 Laszlo Ersek wrote:
The original patches fixing CVE-2023-1076 are incorrect in my opinion. This small series fixes them up; see the individual commit messages for explanation.
I have a very elaborate test procedure demonstrating the problem for both tun and tap; it involves libvirt, qemu, and "crash". I can share that procedure if necessary, but it's indeed quite long (I wrote it originally for our QE team).
Quoted 7 lines
The patches in this series are supposed to "re-fix" CVE-2023-1076; given that said CVE is classified as Low Impact (CVSSv3=5.5), I'm posting this publicly, and not suggesting any embargo. Red Hat Product Security may assign a new CVE number later.
I've tested the patches on top of v6.5-rc4, with "crash" built at commit
c74f375e0ef7.
FTR this was applied yesterday to net. Thanks!