Re: [3.6-rc7] switcheroo race with Intel HDA...

[Daniel J Blueman]

On 9 October 2012 21:04, Takashi Iwai <tiwai@xxxxxxx> wrote:
> At Tue, 9 Oct 2012 19:23:56 +0800,
> Daniel J Blueman wrote:
>> On 9 October 2012 18:07, Takashi Iwai <tiwai@xxxxxxx> wrote:
>> > At Tue, 09 Oct 2012 12:04:08 +0200,
>> > Takashi Iwai wrote:
>> >> At Tue, 9 Oct 2012 00:34:09 +0800,
>> >> Daniel J Blueman wrote:
>> >> > On 8 October 2012 20:58, Takashi Iwai <tiwai@xxxxxxx> wrote:
>> >> > > At Tue, 25 Sep 2012 13:20:05 +0800,
>> >> > > Daniel J Blueman wrote:
>> >> > >> On my Macbook with a discrete Nvidia GPU, there is a race between
>> >> > >> selecting the integrated GPU and putting the discrete GPU into D3 [1],
>> >> > >> reliably causing a kernel oops [2].
>> >> > >>
>> >> > >> Introducing a delay of ~1s between the calls prevents this. When the
>> >> > >> second 'OFF' write path executes, it looks like struct azx at
>> >> > >> card->private_data hasn't yet been allocated yet [3], so there is
>> >> > >> likely some locking missing.
>> >> > >
>> >> > > It's rather pci_get_drvdata() returning NULL (i.e. card is NULL, thus
>> >> > > card->private_data causes Oops).  Could you check the patch like below
>> >> > > and see whether you get a kernel warning (but no Oops) or the problem
>> >> > > gets fixed by shifting the assignment of pci drvdata?
>> >> > [...]
>> >> >
>> >> > Good patching. Calling pci_set_drvdata later prevents the oops in HDA,
>> >> > though we see unexpected 0x0 responses in the response ring buffer
>> >> > [1], which we don't see when there's a >~1.5s delay between IGD and
>> >> > OFF.
>> >>
>> >> If the previous patch fixed, it means that the switching occurred
>> >> during the device was being probed.  Maybe a better approach to
>> >> register the VGA switcheroo after the proper initialization.
>> >>
>> >> The patch below is a revised one.  Please give it a try.
>> >
>> > Also, it's not clear which card spews the spurious response.
>> > Apply the patch below in addition.
>> [...]
>>
>> hda-intel: 0000:01:00.1: spurious response 0x0:0x0, last cmd=0x1f0004
>> $ lspci -s :1:0.1
>> 01:00.1 Audio device: NVIDIA Corporation Device 0e1b (rev ff)
>>
>> It's the NVIDIA device which presumably hasn't completed it's
>> transition to D3 at the time the OFF is executed.
>
> OK, then could you try the patch below on the top of previous two
> patches?

The first IGD switcheroo command fails to switch to the integrated GPU:

# cat /sys/kernel/debug/vgaswitcheroo/switch
0:DIS:+:Pwr:0000:01:00.0
1:IGD: :Pwr:0000:00:02.0
2:DIS-Audio: :Pwr:0000:01:00.1
# echo IGD >/sys/kernel/debug/vgaswitcheroo/switch
vga_switcheroo: client 1 refused switch

I also instrumented snd_hda_lock_devices, but none of the failure
paths are being taken, which would leave inconsistent state, as the
return value isn't checked.

Thanks,
  Daniel
-- 
Daniel J Blueman
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/