net/af_unix: fix a data-race in unix_dgram_poll
authorEric Dumazet <edumazet@google.com>
Thu, 9 Sep 2021 00:00:29 +0000 (17:00 -0700)
committerDavid S. Miller <davem@davemloft.net>
Thu, 9 Sep 2021 09:57:52 +0000 (10:57 +0100)
syzbot reported another data-race in af_unix [1]

Lets change __skb_insert() to use WRITE_ONCE() when changing
skb head qlen.

Also, change unix_dgram_poll() to use lockless version
of unix_recvq_full()

It is verry possible we can switch all/most unix_recvq_full()
to the lockless version, this will be done in a future kernel version.

[1] HEAD commit: 8596e589b787732c8346f0482919e83cc9362db1

BUG: KCSAN: data-race in skb_queue_tail / unix_dgram_poll

write to 0xffff88814eeb24e0 of 4 bytes by task 25815 on cpu 0:
 __skb_insert include/linux/skbuff.h:1938 [inline]
 __skb_queue_before include/linux/skbuff.h:2043 [inline]
 __skb_queue_tail include/linux/skbuff.h:2076 [inline]
 skb_queue_tail+0x80/0xa0 net/core/skbuff.c:3264
 unix_dgram_sendmsg+0xff2/0x1600 net/unix/af_unix.c:1850
 sock_sendmsg_nosec net/socket.c:703 [inline]
 sock_sendmsg net/socket.c:723 [inline]
 ____sys_sendmsg+0x360/0x4d0 net/socket.c:2392
 ___sys_sendmsg net/socket.c:2446 [inline]
 __sys_sendmmsg+0x315/0x4b0 net/socket.c:2532
 __do_sys_sendmmsg net/socket.c:2561 [inline]
 __se_sys_sendmmsg net/socket.c:2558 [inline]
 __x64_sys_sendmmsg+0x53/0x60 net/socket.c:2558
 do_syscall_x64 arch/x86/entry/common.c:50 [inline]
 do_syscall_64+0x3d/0x90 arch/x86/entry/common.c:80
 entry_SYSCALL_64_after_hwframe+0x44/0xae

read to 0xffff88814eeb24e0 of 4 bytes by task 25834 on cpu 1:
 skb_queue_len include/linux/skbuff.h:1869 [inline]
 unix_recvq_full net/unix/af_unix.c:194 [inline]
 unix_dgram_poll+0x2bc/0x3e0 net/unix/af_unix.c:2777
 sock_poll+0x23e/0x260 net/socket.c:1288
 vfs_poll include/linux/poll.h:90 [inline]
 ep_item_poll fs/eventpoll.c:846 [inline]
 ep_send_events fs/eventpoll.c:1683 [inline]
 ep_poll fs/eventpoll.c:1798 [inline]
 do_epoll_wait+0x6ad/0xf00 fs/eventpoll.c:2226
 __do_sys_epoll_wait fs/eventpoll.c:2238 [inline]
 __se_sys_epoll_wait fs/eventpoll.c:2233 [inline]
 __x64_sys_epoll_wait+0xf6/0x120 fs/eventpoll.c:2233
 do_syscall_x64 arch/x86/entry/common.c:50 [inline]
 do_syscall_64+0x3d/0x90 arch/x86/entry/common.c:80
 entry_SYSCALL_64_after_hwframe+0x44/0xae

value changed: 0x0000001b -> 0x00000001

Reported by Kernel Concurrency Sanitizer on:
CPU: 1 PID: 25834 Comm: syz-executor.1 Tainted: G        W         5.14.0-syzkaller #0
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011

Fixes: 86b18aaa2b5b ("skbuff: fix a data race in skb_queue_len()")
Cc: Qian Cai <cai@lca.pw>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/linux/skbuff.h
net/unix/af_unix.c

index 6bdb0db..841e2f0 100644 (file)
@@ -1940,7 +1940,7 @@ static inline void __skb_insert(struct sk_buff *newsk,
        WRITE_ONCE(newsk->prev, prev);
        WRITE_ONCE(next->prev, newsk);
        WRITE_ONCE(prev->next, newsk);
-       list->qlen++;
+       WRITE_ONCE(list->qlen, list->qlen + 1);
 }
 
 static inline void __skb_queue_splice(const struct sk_buff_head *list,
index eb47b9d..92345c9 100644 (file)
@@ -3073,7 +3073,7 @@ static __poll_t unix_dgram_poll(struct file *file, struct socket *sock,
 
                other = unix_peer(sk);
                if (other && unix_peer(other) != sk &&
-                   unix_recvq_full(other) &&
+                   unix_recvq_full_lockless(other) &&
                    unix_dgram_peer_wake_me(sk, other))
                        writable = 0;