Re: switcheroo registration vs switching race...

[Takashi Iwai]

At Tue, 4 Dec 2012 22:46:47 +0800,
Daniel J Blueman wrote:
> 
> On 4 December 2012 21:55, Takashi Iwai <tiwai@xxxxxxx> wrote:
> > At Tue, 04 Dec 2012 14:23:05 +0100,
> > Takashi Iwai wrote:
> >>
> >> At Tue, 4 Dec 2012 20:58:55 +0800,
> >> Daniel J Blueman wrote:
> >> >
> >> > On 4 December 2012 01:10, Takashi Iwai <tiwai@xxxxxxx> wrote:
> >> > > At Tue, 4 Dec 2012 00:50:56 +0800,
> >> > > Daniel J Blueman wrote:
> >> > >>
> >> > >> On 4 December 2012 00:23, Takashi Iwai <tiwai@xxxxxxx> wrote:
> >> > >> > At Mon, 3 Dec 2012 23:08:28 +0800,
> >> > >> > Daniel J Blueman wrote:
> >> > >> >>
> >> > >> >> On 3 December 2012 22:40, Takashi Iwai <tiwai@xxxxxxx> wrote:
> >> > >> >> > At Mon, 3 Dec 2012 22:25:52 +0800,
> >> > >> >> > Daniel J Blueman wrote:
> >> > >> >> >>
> >> > >> >> >> On 3 December 2012 19:17, Takashi Iwai <tiwai@xxxxxxx> wrote:
> >> > >> >> >> > At Wed, 28 Nov 2012 09:45:39 +0100,
> >> > >> >> >> > Takashi Iwai wrote:
> >> > >> >> >> >>
> >> > >> >> >> >> At Wed, 28 Nov 2012 11:45:07 +0800,
> >> > >> >> >> >> Daniel J Blueman wrote:
> >> > >> >> >> >> >
> >> > >> >> >> >> > Hi Seth, Dave, Takashi,
> >> > >> >> >> >> >
> >> > >> >> >> >> > If I power down the unused discrete GPU before lightdm starts by
> >> > >> >> >> >> > fiddling with the sysfs file [1] in the upstart script, I see a race
> >> > >> >> >> >> > manifesting as the discrete GPU's HDA controller timing out to
> >> > >> >> >> >> > commands [2].
> >> > >> >> >> >> >
> >> > >> >> >> >> > Adding some debug, I see that the registered audio devices are put
> >> > >> >> >> >> > into D3 before the GPU is, but it turns out that the discrete (and
> >> > >> >> >> >> > internal) GPU's HDA controller gets registered a bit later, so the
> >> > >> >> >> >> > list is empty. The symptom is since the HDA driver it's talking to
> >> > >> >> >> >> > hardware which is now in D3.
> >> > >> >> >> >> >
> >> > >> >> >> >> > We could add a mutex to nouveau to allow us to wait for the DGPU HDA
> >> > >> >> >> >> > controller, but perhaps this should be solved at a higher level in the
> >> > >> >> >> >> > vgaswitcheroo code; what do you think?
> >> > >> >> >> >>
> >> > >> >> >> >> Maybe it's a side effect for the recent effort to fix another race in
> >> > >> >> >> >> the probe.  A part of them problem is that the registration is done at
> >> > >> >> >> >> the very last of probing.
> >> > >> >> >> >>
> >> > >> >> >> >> Instead of delaying the registration, how about the patch below?
> >> > >> >> >> >
> >> > >> >> >> > Ping.  If this really works, I'd like to queue it for 3.8 merge, at
> >> > >> >> >> > least...
> >> > >> >> >>
> >> > >> >> >> Ping ack; I was trying to find time to understand another race that
> >> > >> >> >> occurs with GPU probing after switching, but is separate from the
> >> > >> >> >> situation before switching, here.
> >> > >> >> >>
> >> > >> >> >> In the context of writing the switch, it looks like struct azx isn't
> >> > >> >> >> allocated by the time azx_vs_set_state accesses it [1,2]; racing with
> >> > >> >> >> azx_codec_create?
> >> > >> >> >
> >> > >> >> > It was allocated, but it wasn't assigned properly in pci drvdata.
> >> > >> >> >
> >> > >> >> > Below is the revised patch.  Just moved pci_set_drvdata() before
> >> > >> >> > register_vga_switcheroo().  Could you retest with it?
> >> > >> >>
> >> > >> >> Superb; this addresses the oops.
> >> > >> >
> >> > >> > OK, I'll queue it to sound tree for 3.8 kernel with Cc to stable.
> >> > >> >
> >> > >> >> ~1 second after the DGPU is put into D3, I still often see "hda-intel:
> >> > >> >> spurious response 0x0:0x0, last cmd=0x170500":
> >> > >> >> http://quora.org/2012/hda-switch-spurious.txt
> >> > >> >
> >> > >> > Hm, it's not clear who triggers these messages.  I'll try to check the
> >> > >> > code paths.
> >> > >> >
> >> > >> >> Presumably this implies the read of the ring-buffer pointer returned
> >> > >> >> 0xffffffff, so the HDA driver understands the pointer to have wrapped
> >> > >> >> and processes the 191 unwritten entries?
> >> > >> >
> >> > >> > Good point.  Actually there is one bug that looks obviously wrong
> >> > >> > (writing 32bit value to CORBWP).  Maybe it has been working just
> >> > >> > because writing CORBRP doesn't influence except for the reset bit.
> >> > >> >
> >> > >> > Reading CORBWP as a byte is OK, but this could be better in a word so
> >> > >> > that we can check 0xffff as invalid.
> >> > >> >
> >> > >> > A test patch is below.  Hopefully this improves the situation...
> >> > >>
> >> > >> I'll check this out tomorrow and also instrument the code to get a
> >> > >> backtrace, since there may still be an underlying race with the
> >> > >> previous patches:
> >> >
> >> > [...]
> >> >
> >> > > That's odd.  The Cirrus one (0000:00:1b.0) must be independent from
> >> > > the vga switcheroo things for Nvidia...
> >> > >
> >> > > The patch below is the revised patch of the first one.  Now the vga
> >> > > switcheroo registration is done before the video controller D3 check,
> >> > > so the race should be smaller.
> >> >
> >> > This patch improves things further of course; a major improvement over
> >> > without. ~15% of the time, I still get the 'spurious response'
> >> > messages in this callpath:
> >>
> >> A possible scenario is that the graphics went in D3 before probing
> >> hd-audio, and the hd-audio continues to initialize the hardware
> >> without knowing the graphics counterpart is disabled.
> >>
> >> There is a code check_hdmi_disabled() in hda_intel.c that checks the
> >> availability of the video driver, and it might be that this doesn't
> >> work as expected...
> >
> > I think I understand this path.  You are setting "OFF", right?
> > This will set the audio client OFF before can_switch() is called.
> > Thus it can be called even before the probing process finished.
> >
> > In short, wait_for_completion() must be put at the beginning of
> > azx_vs_set_state() in addition to azx_vs_can_switch().
> >
> > The revised patch is attached below.  Hopefully this sorts out all
> > races.
> 
> This works great and instead, I get "Cannot lock devices!" sometimes
> (see http://quora.org/2012/hda-switch-lock.txt ).

This means that some apps (likely udev, PulseAudio or whatever) are
accessing sound devices during you turn off the audio / D-GPU.  It
wouldn't be too big trouble, but it's still a bit risky.

We can put the disconnect fops like a real disconnection and restore
in return, but it will need lots of ALSA core code changes, so I won't
follow for now. 

> Feel free to add my
> 'Tested-by: Daniel J Blueman <daniel@xxxxxxxxx>' or ping me if
> anything else to test.

OK, I queued the patch finally to sound git tree for-next branch.

Thanks for your quick testing!


Takashi
--
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/