net: Fix spinlock use in alloc_netdev_mq()
authorIngo Molnar <mingo@elte.hu>
Tue, 4 Aug 2009 21:16:58 +0000 (21:16 +0000)
committerDavid S. Miller <davem@davemloft.net>
Wed, 5 Aug 2009 15:35:11 +0000 (08:35 -0700)
-tip testing found this lockdep warning:

[    2.272010] calling  net_dev_init+0x0/0x164 @ 1
[    2.276033] device class 'net': registering
[    2.280191] INFO: trying to register non-static key.
[    2.284005] the code is fine but needs lockdep annotation.
[    2.284005] turning off the locking correctness validator.
[    2.284005] Pid: 1, comm: swapper Not tainted 2.6.31-rc5-tip #1145
[    2.284005] Call Trace:
[    2.284005]  [<7958eb4e>] ? printk+0xf/0x11
[    2.284005]  [<7904f83c>] __lock_acquire+0x11b/0x622
[    2.284005]  [<7908c9b7>] ? alloc_debug_processing+0xf9/0x144
[    2.284005]  [<7904e2be>] ? mark_held_locks+0x3a/0x52
[    2.284005]  [<7908dbc4>] ? kmem_cache_alloc+0xa8/0x13f
[    2.284005]  [<7904e475>] ? trace_hardirqs_on_caller+0xa2/0xc3
[    2.284005]  [<7904fdf6>] lock_acquire+0xb3/0xd0
[    2.284005]  [<79489678>] ? alloc_netdev_mq+0xf5/0x1ad
[    2.284005]  [<79591514>] _spin_lock_bh+0x2d/0x5d
[    2.284005]  [<79489678>] ? alloc_netdev_mq+0xf5/0x1ad
[    2.284005]  [<79489678>] alloc_netdev_mq+0xf5/0x1ad
[    2.284005]  [<793a38f2>] ? loopback_setup+0x0/0x74
[    2.284005]  [<798eecd0>] loopback_net_init+0x20/0x5d
[    2.284005]  [<79483efb>] register_pernet_device+0x23/0x4b
[    2.284005]  [<798f5c9f>] net_dev_init+0x115/0x164
[    2.284005]  [<7900104f>] do_one_initcall+0x4a/0x11a
[    2.284005]  [<798f5b8a>] ? net_dev_init+0x0/0x164
[    2.284005]  [<79066f6d>] ? register_irq_proc+0x8c/0xa8
[    2.284005]  [<798cc29a>] do_basic_setup+0x42/0x52
[    2.284005]  [<798cc30a>] kernel_init+0x60/0xa1
[    2.284005]  [<798cc2aa>] ? kernel_init+0x0/0xa1
[    2.284005]  [<79003e03>] kernel_thread_helper+0x7/0x10
[    2.284078] device: 'lo': device_add
[    2.288248] initcall net_dev_init+0x0/0x164 returned 0 after 11718 usecs
[    2.292010] calling  neigh_init+0x0/0x66 @ 1
[    2.296010] initcall neigh_init+0x0/0x66 returned 0 after 0 usecs

it's using an zero-initialized spinlock. This is a side-effect of:

        dev_unicast_init(dev);

in alloc_netdev_mq() making use of dev->addr_list_lock.

The device has just been allocated freshly, it's not accessible
anywhere yet so no locking is needed at all - in fact it's wrong
to lock it here (the lock isnt initialized yet).

This bug was introduced via:

| commit a6ac65db2329e7685299666f5f7b6093c7b0f3a0
| Date:   Thu Jul 30 01:06:12 2009 +0000
|
|     net: restore the original spinlock to protect unicast list

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Acked-by: Jiri Pirko <jpirko@redhat.com>
Tested-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/core/dev.c

index 43e61ba..6a94475 100644 (file)
@@ -4007,9 +4007,7 @@ static void dev_unicast_flush(struct net_device *dev)
 
 static void dev_unicast_init(struct net_device *dev)
 {
-       netif_addr_lock_bh(dev);
        __hw_addr_init(&dev->uc);
-       netif_addr_unlock_bh(dev);
 }