Re: BUG: possible deadlock detected! (sound) [Was: 2.6.17-rc5-mm1]

[Takashi Iwai <tiwai@suse.de>]

At Tue, 30 May 2006 13:06:28 +0200,
Arjan van de Ven wrote:
> 
> On Tue, 2006-05-30 at 147 +0159, Jiri Slaby wrote:
> 
> (I've turned your backtrace upside down to show it "chronological")
> 
>  [<c05911e0>] alsa_emu10k1_synth_init+0x22/0x24
>  [<c0333d04>] snd_seq_device_register_driver+0x8f/0xeb
> 
> this one does:
> 
>        mutex_lock(&ops->reg_mutex);
>        ...
>        list_for_each(head, &ops->dev_list) {
>                 struct snd_seq_device *dev = list_entry(head, struct snd_seq_device, list);
>                 init_device(dev, ops);
>        }
>        mutex_unlock(&ops->reg_mutex);
> 
>  [<c0333537>] init_device+0x2c/0x94
>   which calls into the driver
>  [<c0352c39>] snd_emu10k1_synth_new_device+0xe7/0x14e
>  [<c0353f50>] snd_emux_register+0x10d/0x13f
>  [<c0358260>] snd_emux_init_seq_oss+0x35/0x9c
>  [<c0333aa0>] snd_seq_device_new+0x96/0x111
> 
> and this one does
>         mutex_lock(&ops->reg_mutex);
>         list_add_tail(&dev->list, &ops->dev_list);
>         ops->num_devices++;
>         mutex_unlock(&ops->reg_mutex);
> 
> 
> so... on first sight this looks like a real deadlock;
> unless the ALSA folks can tell me why "ops" is always different,
> and what the lock ordering rules between those is...

This ops is a unique object assigned to a different "id" string.

The first snd_seq_device_register_driver() called from emu10k1_synth.c
is the registration for the id "snd-synth-emu10k1".
Then in init_device(), the corresponding devices are initialized, and
one callback registers again another device for OSS sequencer with a
different id "snd-seq-oss" via snd_seq_device_new() inside the lock.
Now it hits the lock-detector but the lock should belong to a
different ops object in practice.

This nested lock may happen only in two drivers, emu10k1-synth and
opl3, and only together with OSS emulation.  Since the OSS emulation
layer don't do active registration from itself, no deadlock should
happen (in theory -- I may oversee something :)


Takashi