openvswitch: fix a possible deadlock and lockdep warning
authorFlavio Leitner <fbl@redhat.com>
Thu, 27 Mar 2014 14:05:34 +0000 (11:05 -0300)
committerDavid S. Miller <davem@davemloft.net>
Fri, 28 Mar 2014 20:41:53 +0000 (16:41 -0400)
There are two problematic situations.

A deadlock can happen when is_percpu is false because it can get
interrupted while holding the spinlock. Then it executes
ovs_flow_stats_update() in softirq context which tries to get
the same lock.

The second sitation is that when is_percpu is true, the code
correctly disables BH but only for the local CPU, so the
following can happen when locking the remote CPU without
disabling BH:

       CPU#0                            CPU#1
  ovs_flow_stats_get()
   stats_read()
 +->spin_lock remote CPU#1        ovs_flow_stats_get()
 |  <interrupted>                  stats_read()
 |  ...                       +-->  spin_lock remote CPU#0
 |                            |     <interrupted>
 |  ovs_flow_stats_update()   |     ...
 |   spin_lock local CPU#0 <--+     ovs_flow_stats_update()
 +---------------------------------- spin_lock local CPU#1

This patch disables BH for both cases fixing the deadlocks.
Acked-by: Jesse Gross <jesse@nicira.com>
=================================
[ INFO: inconsistent lock state ]
3.14.0-rc8-00007-g632b06a #1 Tainted: G          I
---------------------------------
inconsistent {SOFTIRQ-ON-W} -> {IN-SOFTIRQ-W} usage.
swapper/0/0 [HC0[0]:SC1[5]:HE1:SE0] takes:
(&(&cpu_stats->lock)->rlock){+.?...}, at: [<ffffffffa05dd8a1>] ovs_flow_stats_update+0x51/0xd0 [openvswitch]
{SOFTIRQ-ON-W} state was registered at:
[<ffffffff810f973f>] __lock_acquire+0x68f/0x1c40
[<ffffffff810fb4e2>] lock_acquire+0xa2/0x1d0
[<ffffffff817d8d9e>] _raw_spin_lock+0x3e/0x80
[<ffffffffa05dd9e4>] ovs_flow_stats_get+0xc4/0x1e0 [openvswitch]
[<ffffffffa05da855>] ovs_flow_cmd_fill_info+0x185/0x360 [openvswitch]
[<ffffffffa05daf05>] ovs_flow_cmd_build_info.constprop.27+0x55/0x90 [openvswitch]
[<ffffffffa05db41d>] ovs_flow_cmd_new_or_set+0x4dd/0x570 [openvswitch]
[<ffffffff816c245d>] genl_family_rcv_msg+0x1cd/0x3f0
[<ffffffff816c270e>] genl_rcv_msg+0x8e/0xd0
[<ffffffff816c0239>] netlink_rcv_skb+0xa9/0xc0
[<ffffffff816c0798>] genl_rcv+0x28/0x40
[<ffffffff816bf830>] netlink_unicast+0x100/0x1e0
[<ffffffff816bfc57>] netlink_sendmsg+0x347/0x770
[<ffffffff81668e9c>] sock_sendmsg+0x9c/0xe0
[<ffffffff816692d9>] ___sys_sendmsg+0x3a9/0x3c0
[<ffffffff8166a911>] __sys_sendmsg+0x51/0x90
[<ffffffff8166a962>] SyS_sendmsg+0x12/0x20
[<ffffffff817e3ce9>] system_call_fastpath+0x16/0x1b
irq event stamp: 1740726
hardirqs last  enabled at (1740726): [<ffffffff8175d5e0>] ip6_finish_output2+0x4f0/0x840
hardirqs last disabled at (1740725): [<ffffffff8175d59b>] ip6_finish_output2+0x4ab/0x840
softirqs last  enabled at (1740674): [<ffffffff8109be12>] _local_bh_enable+0x22/0x50
softirqs last disabled at (1740675): [<ffffffff8109db05>] irq_exit+0xc5/0xd0

other info that might help us debug this:
 Possible unsafe locking scenario:

       CPU0
       ----
  lock(&(&cpu_stats->lock)->rlock);
  <Interrupt>
    lock(&(&cpu_stats->lock)->rlock);

 *** DEADLOCK ***

5 locks held by swapper/0/0:
 #0:  (((&ifa->dad_timer))){+.-...}, at: [<ffffffff810a7155>] call_timer_fn+0x5/0x320
 #1:  (rcu_read_lock){.+.+..}, at: [<ffffffff81788a55>] mld_sendpack+0x5/0x4a0
 #2:  (rcu_read_lock_bh){.+....}, at: [<ffffffff8175d149>] ip6_finish_output2+0x59/0x840
 #3:  (rcu_read_lock_bh){.+....}, at: [<ffffffff8168ba75>] __dev_queue_xmit+0x5/0x9b0
 #4:  (rcu_read_lock){.+.+..}, at: [<ffffffffa05e41b5>] internal_dev_xmit+0x5/0x110 [openvswitch]

stack backtrace:
CPU: 0 PID: 0 Comm: swapper/0 Tainted: G          I  3.14.0-rc8-00007-g632b06a #1
Hardware name:                  /DX58SO, BIOS SOX5810J.86A.5599.2012.0529.2218 05/29/2012
 0000000000000000 0fcf20709903df0c ffff88042d603808 ffffffff817cfe3c
 ffffffff81c134c0 ffff88042d603858 ffffffff817cb6da 0000000000000005
 ffffffff00000001 ffff880400000000 0000000000000006 ffffffff81c134c0
Call Trace:
 <IRQ>  [<ffffffff817cfe3c>] dump_stack+0x4d/0x66
 [<ffffffff817cb6da>] print_usage_bug+0x1f4/0x205
 [<ffffffff810f7f10>] ? check_usage_backwards+0x180/0x180
 [<ffffffff810f8963>] mark_lock+0x223/0x2b0
 [<ffffffff810f96d3>] __lock_acquire+0x623/0x1c40
 [<ffffffff810f5707>] ? __lock_is_held+0x57/0x80
 [<ffffffffa05e26c6>] ? masked_flow_lookup+0x236/0x250 [openvswitch]
 [<ffffffff810fb4e2>] lock_acquire+0xa2/0x1d0
 [<ffffffffa05dd8a1>] ? ovs_flow_stats_update+0x51/0xd0 [openvswitch]
 [<ffffffff817d8d9e>] _raw_spin_lock+0x3e/0x80
 [<ffffffffa05dd8a1>] ? ovs_flow_stats_update+0x51/0xd0 [openvswitch]
 [<ffffffffa05dd8a1>] ovs_flow_stats_update+0x51/0xd0 [openvswitch]
 [<ffffffffa05dcc64>] ovs_dp_process_received_packet+0x84/0x120 [openvswitch]
 [<ffffffff810f93f7>] ? __lock_acquire+0x347/0x1c40
 [<ffffffffa05e3bea>] ovs_vport_receive+0x2a/0x30 [openvswitch]
 [<ffffffffa05e4218>] internal_dev_xmit+0x68/0x110 [openvswitch]
 [<ffffffffa05e41b5>] ? internal_dev_xmit+0x5/0x110 [openvswitch]
 [<ffffffff8168b4a6>] dev_hard_start_xmit+0x2e6/0x8b0
 [<ffffffff8168be87>] __dev_queue_xmit+0x417/0x9b0
 [<ffffffff8168ba75>] ? __dev_queue_xmit+0x5/0x9b0
 [<ffffffff8175d5e0>] ? ip6_finish_output2+0x4f0/0x840
 [<ffffffff8168c430>] dev_queue_xmit+0x10/0x20
 [<ffffffff8175d641>] ip6_finish_output2+0x551/0x840
 [<ffffffff8176128a>] ? ip6_finish_output+0x9a/0x220
 [<ffffffff8176128a>] ip6_finish_output+0x9a/0x220
 [<ffffffff8176145f>] ip6_output+0x4f/0x1f0
 [<ffffffff81788c29>] mld_sendpack+0x1d9/0x4a0
 [<ffffffff817895b8>] mld_send_initial_cr.part.32+0x88/0xa0
 [<ffffffff817691b0>] ? addrconf_dad_completed+0x220/0x220
 [<ffffffff8178e301>] ipv6_mc_dad_complete+0x31/0x50
 [<ffffffff817690d7>] addrconf_dad_completed+0x147/0x220
 [<ffffffff817691b0>] ? addrconf_dad_completed+0x220/0x220
 [<ffffffff8176934f>] addrconf_dad_timer+0x19f/0x1c0
 [<ffffffff810a71e9>] call_timer_fn+0x99/0x320
 [<ffffffff810a7155>] ? call_timer_fn+0x5/0x320
 [<ffffffff817691b0>] ? addrconf_dad_completed+0x220/0x220
 [<ffffffff810a76c4>] run_timer_softirq+0x254/0x3b0
 [<ffffffff8109d47d>] __do_softirq+0x12d/0x480

Signed-off-by: Flavio Leitner <fbl@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/openvswitch/flow.c

index dda451f4429ca4110d614c6b64b00b59908f0ffa..2998989e76db0a7ccb8e25ef11aa393180956593 100644 (file)
@@ -103,30 +103,24 @@ static void stats_read(struct flow_stats *stats,
 void ovs_flow_stats_get(struct sw_flow *flow, struct ovs_flow_stats *ovs_stats,
                        unsigned long *used, __be16 *tcp_flags)
 {
-       int cpu, cur_cpu;
+       int cpu;
 
        *used = 0;
        *tcp_flags = 0;
        memset(ovs_stats, 0, sizeof(*ovs_stats));
 
+       local_bh_disable();
        if (!flow->stats.is_percpu) {
                stats_read(flow->stats.stat, ovs_stats, used, tcp_flags);
        } else {
-               cur_cpu = get_cpu();
                for_each_possible_cpu(cpu) {
                        struct flow_stats *stats;
 
-                       if (cpu == cur_cpu)
-                               local_bh_disable();
-
                        stats = per_cpu_ptr(flow->stats.cpu_stats, cpu);
                        stats_read(stats, ovs_stats, used, tcp_flags);
-
-                       if (cpu == cur_cpu)
-                               local_bh_enable();
                }
-               put_cpu();
        }
+       local_bh_enable();
 }
 
 static void stats_reset(struct flow_stats *stats)
@@ -141,25 +135,17 @@ static void stats_reset(struct flow_stats *stats)
 
 void ovs_flow_stats_clear(struct sw_flow *flow)
 {
-       int cpu, cur_cpu;
+       int cpu;
 
+       local_bh_disable();
        if (!flow->stats.is_percpu) {
                stats_reset(flow->stats.stat);
        } else {
-               cur_cpu = get_cpu();
-
                for_each_possible_cpu(cpu) {
-
-                       if (cpu == cur_cpu)
-                               local_bh_disable();
-
                        stats_reset(per_cpu_ptr(flow->stats.cpu_stats, cpu));
-
-                       if (cpu == cur_cpu)
-                               local_bh_enable();
                }
-               put_cpu();
        }
+       local_bh_enable();
 }
 
 static int check_header(struct sk_buff *skb, int len)