tipc: fix potential deadlock when all links are reset
authorYing Xue <ying.xue@windriver.com>
Thu, 26 Mar 2015 10:10:23 +0000 (18:10 +0800)
committerDavid S. Miller <davem@davemloft.net>
Sun, 29 Mar 2015 19:40:27 +0000 (12:40 -0700)
commitb952b2befb6f6b009e91f087285b9a0a6beb1cc8
tree39723b22169b63f4f6b81fdb60ad618053271879
parentfaadb05f4b7f11b3063a20d7fc5afcbf1124dbeb
tipc: fix potential deadlock when all links are reset

[   60.988363] ======================================================
[   60.988754] [ INFO: possible circular locking dependency detected ]
[   60.989152] 3.19.0+ #194 Not tainted
[   60.989377] -------------------------------------------------------
[   60.989781] swapper/3/0 is trying to acquire lock:
[   60.990079]  (&(&n_ptr->lock)->rlock){+.-...}, at: [<ffffffffa0006dca>] tipc_link_retransmit+0x1aa/0x240 [tipc]
[   60.990743]
[   60.990743] but task is already holding lock:
[   60.991106]  (&(&bclink->lock)->rlock){+.-...}, at: [<ffffffffa00004be>] tipc_bclink_lock+0x8e/0xa0 [tipc]
[   60.991738]
[   60.991738] which lock already depends on the new lock.
[   60.991738]
[   60.992174]
[   60.992174] the existing dependency chain (in reverse order) is:
[   60.992174]
-> #1 (&(&bclink->lock)->rlock){+.-...}:
[   60.992174]        [<ffffffff810a9c0c>] lock_acquire+0x9c/0x140
[   60.992174]        [<ffffffff8179c41f>] _raw_spin_lock_bh+0x3f/0x50
[   60.992174]        [<ffffffffa00004be>] tipc_bclink_lock+0x8e/0xa0 [tipc]
[   60.992174]        [<ffffffffa0000f57>] tipc_bclink_add_node+0x97/0xf0 [tipc]
[   60.992174]        [<ffffffffa0011815>] tipc_node_link_up+0xf5/0x110 [tipc]
[   60.992174]        [<ffffffffa0007783>] link_state_event+0x2b3/0x4f0 [tipc]
[   60.992174]        [<ffffffffa00193c0>] tipc_link_proto_rcv+0x24c/0x418 [tipc]
[   60.992174]        [<ffffffffa0008857>] tipc_rcv+0x827/0xac0 [tipc]
[   60.992174]        [<ffffffffa0002ca3>] tipc_l2_rcv_msg+0x73/0xd0 [tipc]
[   60.992174]        [<ffffffff81646e66>] __netif_receive_skb_core+0x746/0x980
[   60.992174]        [<ffffffff816470c1>] __netif_receive_skb+0x21/0x70
[   60.992174]        [<ffffffff81647295>] netif_receive_skb_internal+0x35/0x130
[   60.992174]        [<ffffffff81648218>] napi_gro_receive+0x158/0x1d0
[   60.992174]        [<ffffffff81559e05>] e1000_clean_rx_irq+0x155/0x490
[   60.992174]        [<ffffffff8155c1b7>] e1000_clean+0x267/0x990
[   60.992174]        [<ffffffff81647b60>] net_rx_action+0x150/0x360
[   60.992174]        [<ffffffff8105ec43>] __do_softirq+0x123/0x360
[   60.992174]        [<ffffffff8105f12e>] irq_exit+0x8e/0xb0
[   60.992174]        [<ffffffff8179f9f5>] do_IRQ+0x65/0x110
[   60.992174]        [<ffffffff8179da6f>] ret_from_intr+0x0/0x13
[   60.992174]        [<ffffffff8100de9f>] arch_cpu_idle+0xf/0x20
[   60.992174]        [<ffffffff8109dfa6>] cpu_startup_entry+0x2f6/0x3f0
[   60.992174]        [<ffffffff81033cda>] start_secondary+0x13a/0x150
[   60.992174]
-> #0 (&(&n_ptr->lock)->rlock){+.-...}:
[   60.992174]        [<ffffffff810a8f7d>] __lock_acquire+0x163d/0x1ca0
[   60.992174]        [<ffffffff810a9c0c>] lock_acquire+0x9c/0x140
[   60.992174]        [<ffffffff8179c41f>] _raw_spin_lock_bh+0x3f/0x50
[   60.992174]        [<ffffffffa0006dca>] tipc_link_retransmit+0x1aa/0x240 [tipc]
[   60.992174]        [<ffffffffa0001e11>] tipc_bclink_rcv+0x611/0x640 [tipc]
[   60.992174]        [<ffffffffa0008646>] tipc_rcv+0x616/0xac0 [tipc]
[   60.992174]        [<ffffffffa0002ca3>] tipc_l2_rcv_msg+0x73/0xd0 [tipc]
[   60.992174]        [<ffffffff81646e66>] __netif_receive_skb_core+0x746/0x980
[   60.992174]        [<ffffffff816470c1>] __netif_receive_skb+0x21/0x70
[   60.992174]        [<ffffffff81647295>] netif_receive_skb_internal+0x35/0x130
[   60.992174]        [<ffffffff81648218>] napi_gro_receive+0x158/0x1d0
[   60.992174]        [<ffffffff81559e05>] e1000_clean_rx_irq+0x155/0x490
[   60.992174]        [<ffffffff8155c1b7>] e1000_clean+0x267/0x990
[   60.992174]        [<ffffffff81647b60>] net_rx_action+0x150/0x360
[   60.992174]        [<ffffffff8105ec43>] __do_softirq+0x123/0x360
[   60.992174]        [<ffffffff8105f12e>] irq_exit+0x8e/0xb0
[   60.992174]        [<ffffffff8179f9f5>] do_IRQ+0x65/0x110
[   60.992174]        [<ffffffff8179da6f>] ret_from_intr+0x0/0x13
[   60.992174]        [<ffffffff8100de9f>] arch_cpu_idle+0xf/0x20
[   60.992174]        [<ffffffff8109dfa6>] cpu_startup_entry+0x2f6/0x3f0
[   60.992174]        [<ffffffff81033cda>] start_secondary+0x13a/0x150
[   60.992174]
[   60.992174] other info that might help us debug this:
[   60.992174]
[   60.992174]  Possible unsafe locking scenario:
[   60.992174]
[   60.992174]        CPU0                    CPU1
[   60.992174]        ----                    ----
[   60.992174]   lock(&(&bclink->lock)->rlock);
[   60.992174]                                lock(&(&n_ptr->lock)->rlock);
[   60.992174]                                lock(&(&bclink->lock)->rlock);
[   60.992174]   lock(&(&n_ptr->lock)->rlock);
[   60.992174]
[   60.992174]  *** DEADLOCK ***
[   60.992174]
[   60.992174] 3 locks held by swapper/3/0:
[   60.992174]  #0:  (rcu_read_lock){......}, at: [<ffffffff81646791>] __netif_receive_skb_core+0x71/0x980
[   60.992174]  #1:  (rcu_read_lock){......}, at: [<ffffffffa0002c35>] tipc_l2_rcv_msg+0x5/0xd0 [tipc]
[   60.992174]  #2:  (&(&bclink->lock)->rlock){+.-...}, at: [<ffffffffa00004be>] tipc_bclink_lock+0x8e/0xa0 [tipc]
[   60.992174]

The correct the sequence of grabbing n_ptr->lock and bclink->lock
should be that the former is first held and the latter is then taken,
which exactly happened on CPU1. But especially when the retransmission
of broadcast link is failed, bclink->lock is first held in
tipc_bclink_rcv(), and n_ptr->lock is taken in link_retransmit_failure()
called by tipc_link_retransmit() subsequently, which is demonstrated on
CPU0. As a result, deadlock occurs.

If the order of holding the two locks happening on CPU0 is reversed, the
deadlock risk will be relieved. Therefore, the node lock taken in
link_retransmit_failure() originally is moved to tipc_bclink_rcv()
so that it's obtained before bclink lock. But the precondition of
the adjustment of node lock is that responding to bclink reset event
must be moved from tipc_bclink_unlock() to tipc_node_unlock().

Reviewed-by: Erik Hugne <erik.hugne@ericsson.com>
Signed-off-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/tipc/bcast.c
net/tipc/bcast.h
net/tipc/link.c
net/tipc/node.c
net/tipc/node.h