tipc: fix lockdep warning on address assignment
authorNeil Horman <nhorman@tuxdriver.com>
Tue, 16 Mar 2010 08:14:33 +0000 (08:14 +0000)
committerDavid S. Miller <davem@davemloft.net>
Tue, 16 Mar 2010 21:15:45 +0000 (14:15 -0700)
commita2f46ee1ba5ee249ce2ca1ee7a7a0ac46529fb4f
treea791769005885ab26a3353b70940c635b63b3e8d
parentc3259c8a7060d480e8eb2166da0a99d6879146b4
tipc: fix lockdep warning on address assignment

So in the forward porting of various tipc packages, I was constantly
getting this lockdep warning everytime I used tipc-config to set a network
address for the protocol:

[ INFO: possible circular locking dependency detected ]
2.6.33 #1
tipc-config/1326 is trying to acquire lock:
(ref_table_lock){+.-...}, at: [<ffffffffa0315148>] tipc_ref_discard+0x53/0xd4 [tipc]

but task is already holding lock:
(&(&entry->lock)->rlock#2){+.-...}, at: [<ffffffffa03150d5>] tipc_ref_lock+0x43/0x63 [tipc]

which lock already depends on the new lock.

the existing dependency chain (in reverse order) is:

-> #1 (&(&entry->lock)->rlock#2){+.-...}:
[<ffffffff8107b508>] __lock_acquire+0xb67/0xd0f
[<ffffffff8107b78c>] lock_acquire+0xdc/0x102
[<ffffffff8145471e>] _raw_spin_lock_bh+0x3b/0x6e
[<ffffffffa03152b1>] tipc_ref_acquire+0xe8/0x11b [tipc]
[<ffffffffa031433f>] tipc_createport_raw+0x78/0x1b9 [tipc]
[<ffffffffa031450b>] tipc_createport+0x8b/0x125 [tipc]
[<ffffffffa030f221>] tipc_subscr_start+0xce/0x126 [tipc]
[<ffffffffa0308fb2>] process_signal_queue+0x47/0x7d [tipc]
[<ffffffff81053e0c>] tasklet_action+0x8c/0xf4
[<ffffffff81054bd8>] __do_softirq+0xf8/0x1cd
[<ffffffff8100aadc>] call_softirq+0x1c/0x30
[<ffffffff810549f4>] _local_bh_enable_ip+0xb8/0xd7
[<ffffffff81054a21>] local_bh_enable_ip+0xe/0x10
[<ffffffff81454d31>] _raw_spin_unlock_bh+0x34/0x39
[<ffffffffa0308eb8>] spin_unlock_bh.clone.0+0x15/0x17 [tipc]
[<ffffffffa0308f47>] tipc_k_signal+0x8d/0xb1 [tipc]
[<ffffffffa0308dd9>] tipc_core_start+0x8a/0xad [tipc]
[<ffffffffa01b1087>] 0xffffffffa01b1087
[<ffffffff8100207d>] do_one_initcall+0x72/0x18a
[<ffffffff810872fb>] sys_init_module+0xd8/0x23a
[<ffffffff81009b42>] system_call_fastpath+0x16/0x1b

-> #0 (ref_table_lock){+.-...}:
[<ffffffff8107b3b2>] __lock_acquire+0xa11/0xd0f
[<ffffffff8107b78c>] lock_acquire+0xdc/0x102
[<ffffffff81454836>] _raw_write_lock_bh+0x3b/0x6e
[<ffffffffa0315148>] tipc_ref_discard+0x53/0xd4 [tipc]
[<ffffffffa03141ee>] tipc_deleteport+0x40/0x119 [tipc]
[<ffffffffa0316e35>] release+0xeb/0x137 [tipc]
[<ffffffff8139dbf4>] sock_release+0x1f/0x6f
[<ffffffff8139dc6b>] sock_close+0x27/0x2b
[<ffffffff811116f6>] __fput+0x12a/0x1df
[<ffffffff811117c5>] fput+0x1a/0x1c
[<ffffffff8110e49b>] filp_close+0x68/0x72
[<ffffffff8110e552>] sys_close+0xad/0xe7
[<ffffffff81009b42>] system_call_fastpath+0x16/0x1b

Finally decided I should fix this.  Its a straightforward inversion,
tipc_ref_acquire takes two locks in this order:
ref_table_lock
entry->lock

while tipc_deleteport takes them in this order:
entry->lock (via tipc_port_lock())
ref_table_lock (via tipc_ref_discard())

when the same entry is referenced, we get the above warning.  The fix is equally
straightforward.  Theres no real relation between the entry->lock and the
ref_table_lock (they just are needed at the same time), so move the entry->lock
aquisition in tipc_ref_acquire down, after we unlock ref_table_lock (this is
safe since the ref_table_lock guards changes to the reference table, and we've
already claimed a slot there.  I've tested the below fix and confirmed that it
clears up the lockdep issue

Signed-off-by: Neil Horman <nhorman@tuxdriver.com>
CC: Allan Stephens <allan.stephens@windriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/tipc/ref.c