Re: mlx5: net_device.addr_list_lock usage before initialization

[Saeed Mahameed]

On Tue, Dec 13, 2016 at 3:22 PM, Sebastian Ott
 wrote:
> Hi,
>
> I ran into the following lockdep complaint:
>
> [7.059561] INFO: trying to register non-static key.
> [7.059566] the code is fine but needs lockdep annotation.
> [7.059570] turning off the locking correctness validator.
> [7.059579] CPU: 6 PID: 6 Comm: kworker/u32:0 Not tainted 
> 4.9.0-02683-g784243e-dirty #77
> [7.059582] Hardware name: IBM  2964 N96  704  
> (LPAR)
> [7.061260] Workqueue: mlx5e mlx5e_set_rx_mode_work [mlx5_core]
> [7.061268] Stack:
> [7.061270]f95739c0 f9573a50 0003 
> 
> [7.061278]f9573af0 f9573a68 f9573a68 
> 0020
> [7.061286] 0020 000a 
> 000a
> [7.061294]000c f9573ab8  
> 
> [7.061301]008a1038 00112a50 f9573a50 
> f9573aa8
> [7.061314] Call Trace:
> [7.061321] ([<0011292a>] show_trace+0x8a/0xe0)
> [7.061327]  [<00112a00>] show_stack+0x80/0xd8
> [7.061334]  [<005cdce6>] dump_stack+0x96/0xd8
> [7.061338]  [<001ae352>] register_lock_class+0x1d2/0x530
> [7.061341]  [<001b33f6>] __lock_acquire+0xfe/0x7d8
> [7.061345]  [<001b4394>] lock_acquire+0x30c/0x358
> [7.061352]  [<0089454c>] _raw_spin_lock_bh+0x64/0xa0
> [7.062171]  [<03ff81465858>] mlx5e_set_rx_mode_work+0x248/0x490 
> [mlx5_core]
> [7.062178]  [<00163864>] process_one_work+0x41c/0x830
> [7.062181]  [<00163f2c>] worker_thread+0x2b4/0x478
> [7.062186]  [<0016c46c>] kthread+0x15c/0x170
> [7.062190]  [<00895a52>] kernel_thread_starter+0x6/0xc
> [7.062193]  [<00895a4c>] kernel_thread_starter+0x0/0xc
> [7.062196] INFO: lockdep is turned off.
>
> The problematic lock is net_device.addr_list_lock whose usage is
> asynchronously triggered by:
>
> mlx5e_add -> mlx5e_attach -> mlx5e_attach_netdev -> mlx5e_nic_enable
> [workq] mlx5e_set_rx_mode_work -> mlx5e_handle_netdev_addr -> 
> mlx5e_sync_netdev_addr
>
> Initialization of this lock is triggered by:
> mlx5e_add -> register_netdev
>
> ...after the call to mlx5e_attach which is obviously racy.
>

Thanks Sebastian for the report,

indeed there is an issue, I wonder why the net_device.addr_list_lock
is initialized so late (at register_netdevice) IMHO it should be
initialized at
alloc_netdev_mqs->dev_addr_init
where all the other net_device fields are initialized!

We will handle this.

Thanks,
Saeed.

mlx5: net_device.addr_list_lock usage before initialization

[Sebastian Ott]

Hi,

I ran into the following lockdep complaint:

[7.059561] INFO: trying to register non-static key.
[7.059566] the code is fine but needs lockdep annotation.
[7.059570] turning off the locking correctness validator.
[7.059579] CPU: 6 PID: 6 Comm: kworker/u32:0 Not tainted 
4.9.0-02683-g784243e-dirty #77
[7.059582] Hardware name: IBM  2964 N96  704
  (LPAR)
[7.061260] Workqueue: mlx5e mlx5e_set_rx_mode_work [mlx5_core]
[7.061268] Stack:
[7.061270]f95739c0 f9573a50 0003 

[7.061278]f9573af0 f9573a68 f9573a68 
0020
[7.061286] 0020 000a 
000a
[7.061294]000c f9573ab8  

[7.061301]008a1038 00112a50 f9573a50 
f9573aa8
[7.061314] Call Trace:
[7.061321] ([<0011292a>] show_trace+0x8a/0xe0)
[7.061327]  [<00112a00>] show_stack+0x80/0xd8
[7.061334]  [<005cdce6>] dump_stack+0x96/0xd8
[7.061338]  [<001ae352>] register_lock_class+0x1d2/0x530
[7.061341]  [<001b33f6>] __lock_acquire+0xfe/0x7d8
[7.061345]  [<001b4394>] lock_acquire+0x30c/0x358
[7.061352]  [<0089454c>] _raw_spin_lock_bh+0x64/0xa0
[7.062171]  [<03ff81465858>] mlx5e_set_rx_mode_work+0x248/0x490 
[mlx5_core]
[7.062178]  [<00163864>] process_one_work+0x41c/0x830
[7.062181]  [<00163f2c>] worker_thread+0x2b4/0x478
[7.062186]  [<0016c46c>] kthread+0x15c/0x170
[7.062190]  [<00895a52>] kernel_thread_starter+0x6/0xc
[7.062193]  [<00895a4c>] kernel_thread_starter+0x0/0xc
[7.062196] INFO: lockdep is turned off.

The problematic lock is net_device.addr_list_lock whose usage is
asynchronously triggered by:

mlx5e_add -> mlx5e_attach -> mlx5e_attach_netdev -> mlx5e_nic_enable 
[workq] mlx5e_set_rx_mode_work -> mlx5e_handle_netdev_addr -> 
mlx5e_sync_netdev_addr

Initialization of this lock is triggered by:
mlx5e_add -> register_netdev

...after the call to mlx5e_attach which is obviously racy.

Regards,
Sebastian