Re: mmotm 2009-08-04-14-22 uploaded

On Wed, 5 Aug 2009 14:39:46 +0800 Dave Young [off-list ref] wrote:

Hi andrew,

I see following lockdep warning with this release:

[    0.474144] INFO: trying to register non-static key.
[    0.474144] the code is fine but needs lockdep annotation.
[    0.474144] turning off the locking correctness validator.
[    0.474144] Pid: 1, comm: swapper Not tainted 2.6.31-rc5-mm1 #7
[    0.474144] Call Trace:
[    0.474144]  [<c1047f1e>] register_lock_class+0x58/0x241
[    0.474144]  [<c1049ab1>] __lock_acquire+0xac/0xb73
[    0.474144]  [<c1076eb5>] ? __alloc_pages_nodemask+0xe2/0x483
[    0.474144]  [<c1048b64>] ? mark_lock+0x1e/0x1c7
[    0.474144]  [<c1048b64>] ? mark_lock+0x1e/0x1c7
[    0.474144]  [<c1048d50>] ? mark_held_locks+0x43/0x5b
[    0.474144]  [<c10940a6>] ? kmem_cache_alloc+0xac/0x11b
[    0.474144]  [<c104a615>] lock_acquire+0x9d/0xc0
[    0.474144]  [<c12b8b96>] ? netif_addr_lock_bh+0xd/0xf
[    0.474144]  [<c1330feb>] _spin_lock_bh+0x20/0x2f
[    0.474144]  [<c12b8b96>] ? netif_addr_lock_bh+0xd/0xf
[    0.474144]  [<c12b8b96>] netif_addr_lock_bh+0xd/0xf
[    0.474144]  [<c12bc3c3>] alloc_netdev_mq+0xf9/0x1a5
[    0.474144]  [<c121f016>] ? loopback_setup+0x0/0x74
[    0.474144]  [<c1578d49>] loopback_net_init+0x20/0x5d
[    0.474144]  [<c12b7907>] register_pernet_operations+0x13/0x15
[    0.474144]  [<c12b7970>] register_pernet_device+0x1f/0x47
[    0.474144]  [<c157ee8d>] net_dev_init+0xfe/0x14d
[    0.474144]  [<c1001137>] do_one_initcall+0x4a/0x11a
[    0.474144]  [<c157ed8f>] ? net_dev_init+0x0/0x14d
[    0.474144]  [<c1067e00>] ? register_irq_proc+0x64/0xa8
[    0.474144]  [<c1067e97>] ? init_irq_proc+0x53/0x60
[    0.474144]  [<c1557535>] kernel_init+0x129/0x17a
[    0.474144]  [<c155740c>] ? kernel_init+0x0/0x17a
[    0.474144]  [<c1003d47>] kernel_thread_helper+0x7/0x10

At a guess I'd say that alloc_netdev_mq()->dev_unicast_init() is doing
netif_addr_lock_bh()->spin_lock_bh(&dev->addr_list_lock) prior to
initialising add_list_lock.

Something like this might shut it up:

--- a/net/core/dev.c~a
+++ a/net/core/dev.c

@@ -5111,7 +5111,7 @@ struct net_device *alloc_netdev_mq(int s
 	if (dev_addr_init(dev))
 		goto free_tx;
 
-	dev_unicast_init(dev);
+	__hw_addr_init(&dev->uc);
 
 	dev_net_set(dev, &init_net);
 

but it'd be better to intialise this thing earlier like:

--- a/net/core/dev.c~a
+++ a/net/core/dev.c

@@ -4730,8 +4730,6 @@ int register_netdevice(struct net_device
 	BUG_ON(dev->reg_state != NETREG_UNINITIALIZED);
 	BUG_ON(!net);
 
-	spin_lock_init(&dev->addr_list_lock);
-	netdev_set_addr_lockdep_class(dev);
 	netdev_init_queue_locks(dev);
 
 	dev->iflink = -1;

@@ -5107,6 +5105,8 @@ struct net_device *alloc_netdev_mq(int s
 
 	dev = PTR_ALIGN(p, NETDEV_ALIGN);
 	dev->padded = (char *)dev - (char *)p;
+	spin_lock_init(&dev->addr_list_lock);
+	netdev_set_addr_lockdep_class(dev);
 
 	if (dev_addr_init(dev))
 		goto free_tx;

_

but that might break register_netdevice() for netdevs which were
allocated via other means, dunno.

I would be pointing fingers at

: commit 31278e71471399beaff9280737e52b47db4dc345
: Author:     Jiri Pirko [off-list ref]
: AuthorDate: Wed Jun 17 01:12:19 2009 +0000
: Commit:     David S. Miller [off-list ref]
: CommitDate: Thu Jun 18 00:29:08 2009 -0700
: 
:     net: group address list and its count
    
and politely suggesting that net developers enable lockdep when testing :)