Re: [Bug #11380] lockdep warning: cpu_add_remove_lock at:cpu_maps_update_begin+0x14/0x16

[Oleg Nesterov <oleg-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>]

On 11/11, KOSAKI Motohiro wrote:
>
> it seem simple ABBA lock, right?
>
> -> #4 (&dev->mutex){--..}:
>        [<c0160f87>] validate_chain+0x831/0xaa2
>        [<c0161872>] __lock_acquire+0x67a/0x6e0
>        [<c0161933>] lock_acquire+0x5b/0x81
>        [<c0a660e4>] mutex_lock_interruptible_nested+0xde/0x2f8
>        [<c0782d02>] input_register_handle+0x26/0x7a			dev->mutex
                                                                        ^^^^^^^^^^
>        [<c04a62c9>] kbd_connect+0x64/0x8d
>        [<c0782842>] input_attach_handler+0x38/0x6b
>        [<c0784216>] input_register_handler+0x74/0xc3			input_mutex
>        [<c0f54e4b>] kbd_init+0x66/0x91
>        [<c0f54f7b>] vty_init+0xce/0xd7
>        [<c0f54952>] tty_init+0x193/0x197
>        [<c010112a>] do_one_initcall+0x42/0x133
>        [<c0f2d5cb>] kernel_init+0x16e/0x1d5
>        [<c0117c03>] kernel_thread_helper+0x7/0x10
>        [<ffffffff>] 0xffffffff
>
> -> #3 (input_mutex){--..}:
>        [<c0160f87>] validate_chain+0x831/0xaa2
>        [<c0161872>] __lock_acquire+0x67a/0x6e0
>        [<c0161933>] lock_acquire+0x5b/0x81
>        [<c0a660e4>] mutex_lock_interruptible_nested+0xde/0x2f8
>        [<c0783fbe>] input_register_device+0xff/0x17f		input_mutex
>        [<c048a889>] acpi_button_add+0x31e/0x429
>        [<c04889f4>] acpi_device_probe+0x43/0xde
>        [<c052c67f>] driver_probe_device+0xa5/0x120
>        [<c052c73c>] __driver_attach+0x42/0x64			dev->sem
                                                                ^^^^^^^^
input_dev->mutex != device->sem

> ...
>        [<c0f2d201>] do_async_initcalls+0x1a/0x2a
>        [<c0150eec>] run_workqueue+0xc3/0x193			
>        [<c015195d>] worker_thread+0xbb/0xc7
>        [<c0153e2a>] kthread+0x40/0x66

What is the kernel version, btw? I can't find do_async_initcalls
in 2.6.27 or 2.6.28.

Anyway, this really looks like lockdep bug to me. Even if we really
have the circular dependency (will try to grep more) I can't understand
why lockdep claims that polldev_mutex depends on cpu_add_remove_lock.

Oleg.