On Tue, May 15, 2018 at 8:18 AM, Andrei Vagin <ava...@virtuozzo.com> wrote: >> >> >> Hello, >> >> >> >> >> >> syzbot found the following crash on: >> >> >> >> >> >> HEAD commit: c1c07416cdd4 Merge tag 'kbuild-fixes-v4.17' of >> >> >> git://git.k.. >> >> >> git tree: upstream >> >> >> console output: >> >> >> https://syzkaller.appspot.com/x/log.txt?x=12164c97800000 >> >> >> kernel config: >> >> >> https://syzkaller.appspot.com/x/.config?x=5a1dc06635c10d27 >> >> >> dashboard link: >> >> >> https://syzkaller.appspot.com/bug?extid=c1872be62e587eae9669 >> >> >> compiler: gcc (GCC) 8.0.1 20180413 (experimental) >> >> >> userspace arch: i386 >> >> >> >> >> >> Unfortunately, I don't have any reproducer for this crash yet. >> >> >> >> >> >> IMPORTANT: if you fix the bug, please add the following tag to the >> >> >> commit: >> >> >> Reported-by: syzbot+c1872be62e587eae9...@syzkaller.appspotmail.com >> >> >> >> >> >> >> >> >> ====================================================== >> >> >> WARNING: possible circular locking dependency detected >> >> >> 4.17.0-rc3+ #59 Not tainted >> >> >> ------------------------------------------------------ >> >> >> syz-executor1/25282 is trying to acquire lock: >> >> >> 000000004fddf743 (&(&u->lock)->rlock/1){+.+.}, at: sk_diag_dump_icons >> >> >> net/unix/diag.c:82 [inline] >> >> >> 000000004fddf743 (&(&u->lock)->rlock/1){+.+.}, at: >> >> >> sk_diag_fill.isra.5+0xa43/0x10d0 net/unix/diag.c:144 >> >> >> >> >> >> but task is already holding lock: >> >> >> 00000000b6895645 (rlock-AF_UNIX){+.+.}, at: spin_lock >> >> >> include/linux/spinlock.h:310 [inline] >> >> >> 00000000b6895645 (rlock-AF_UNIX){+.+.}, at: sk_diag_dump_icons >> >> >> net/unix/diag.c:64 [inline] >> >> >> 00000000b6895645 (rlock-AF_UNIX){+.+.}, at: >> >> >> sk_diag_fill.isra.5+0x94e/0x10d0 >> >> >> net/unix/diag.c:144 >> >> >> >> >> >> which lock already depends on the new lock. >> >> > >> >> > In the code, we have a comment which explains why it is safe to take >> >> > this lock >> >> > >> >> > /* >> >> > * The state lock is outer for the same sk's >> >> > * queue lock. With the other's queue locked it's >> >> > * OK to lock the state. >> >> > */ >> >> > unix_state_lock_nested(req); >> >> > >> >> > It is a question how to explain this to lockdep. >> >> >> >> Do I understand it correctly that (&u->lock)->rlock associated with >> >> AF_UNIX is locked under rlock-AF_UNIX, and then rlock-AF_UNIX is >> >> locked under (&u->lock)->rlock associated with AF_NETLINK? If so, I >> >> think we need to split (&u->lock)->rlock by family too, so that we >> >> have u->lock-AF_UNIX and u->lock-AF_NETLINK. >> > >> > I think here is another problem. lockdep woried about >> > sk->sk_receive_queue vs unix_sk(s)->lock. >> > >> > sk_diag_dump_icons() takes sk->sk_receive_queue and then >> > unix_sk(s)->lock. >> > >> > unix_dgram_sendmsg takes unix_sk(sk)->lock and then sk->sk_receive_queue. >> > >> > sk_diag_dump_icons() takes locks for two different sockets, but >> > unix_dgram_sendmsg() takes locks for one socket. >> > >> > sk_diag_dump_icons >> > if (sk->sk_state == TCP_LISTEN) { >> > spin_lock(&sk->sk_receive_queue.lock); >> > skb_queue_walk(&sk->sk_receive_queue, skb) { >> > unix_state_lock_nested(req); >> > spin_lock_nested(&unix_sk(s)->lock, >> > >> > >> > unix_dgram_sendmsg >> > unix_state_lock(other) >> > spin_lock(&unix_sk(s)->lock) >> > skb_queue_tail(&other->sk_receive_queue, skb); >> > spin_lock_irqsave(&list->lock, flags); >> >> >> Do you mean the following? >> There is socket 1 with state lock (S1) and queue lock (Q2), and socket >> 2 with state lock (S2) and queue lock (Q2). unix_dgram_sendmsg lock >> S1->Q1. And sk_diag_dump_icons locks Q1->S2. >> If yes, then this looks pretty much as deadlock. Consider that 2 >> unix_dgram_sendmsg in 2 different threads lock S1 and S2 respectively. >> Now 2 sk_diag_dump_icons in 2 different threads lock Q1 and Q2 >> respectively. Now sk_diag_dump_icons want to lock S's, and >> unix_dgram_sendmsg want to lock Q's. Nobody can proceed. > > Q1 and S1 belongs to a listen socket, so they can't be taken from > unix_dgram_sendmsg().
Should we then split Q1/S1 for listening and data sockets? I don't know it lockdep allows changing lock class on the fly, though. Always wondered if there was a single reason to mix listening and data sockets into a single thing on API level...
