Re: Crash in acpi_ns_validate_handle triggered by soundwire on Linux 5.10

["Rafael J. Wysocki" <rafael@kernel.org>]

On Thu, Jan 28, 2021 at 2:25 PM Marcin Ślusarz <marcin.slusarz@gmail.com> wrote:
>
> śr., 27 sty 2021 o 23:02 Pierre-Louis Bossart
> <pierre-louis.bossart@linux.intel.com> napisał(a):
> > On 1/27/21 1:18 PM, Marcin Ślusarz wrote:
> > > śr., 27 sty 2021 o 18:28 Pierre-Louis Bossart
> > > <pierre-louis.bossart@linux.intel.com> napisał(a):
> > >>> Weird, I can't reproduce this problem with my self-compiled kernel :/
> > >>> I don't even see soundwire modules loaded in. Manually loading them of course
> > >>> doesn't do much.
> > >>>
> > >>> Previously I could boot into the "faulty" kernel by using "recovery mode", but
> > >>> I can't do that anymore - it crashes too.
> > >>>
> > >>> Maybe there's some kind of race and this bug depends on some specific
> > >>> ordering of events?
> > >>
> > >> missing Kconfig?
> > >> You need CONFIG_SOUNDWIRE and CONFIG_SND_SOC_SOF_INTEL_SOUNDWIRE
> > >> selected to enter this sdw_intel_acpi_scan() routine.
> > >
> > > It was a PEBKAC, but a slightly different one. I won't bore you with
> > > (embarrassing) details ;).
> > >
> > > I reproduced the problem, tested both your and Rafael's patches
> > > and the kernel still crashes, with the same stack trace.
> > > (Yes, I'm sure I booted the right kernel :)
> > >
> > > Why "recovery mode" stopped working (or worked previously) is still a mystery.
> > >
> >
> > Thanks Marcin for the information. If you have a consistent failure
> > that's better to some extent.
> >
> > Maybe a bit of explanation of what this routine tries to do:
> > when SoundWire is enabled in a system, we need to have the following
> > pattern in the DSDT:
> >
> >      Scope (_SB.PCI0)
> >      {
> >          Device (HDAS)
> >          {
> >              Name (_ADR, 0x001F0003)  // _ADR: Address
> >          }
> >
> >
> >          Scope (HDAS)
> >          {
> >              Device (SNDW)
> >              {
> >                  Name (_ADR, 0x40000000)  // _ADR: Address
> >
> > The only thing the code does is to walk through the children and check
> > if the valid _ADR 0x40000000 is found.
> >
> > You don't have SoundWire in your device so there should not be any
> > children found. I don't see anything in the DSDT that looks like
> > _SB.PCI0.HDAS.<something>, so in theory we should not even enter the
> > callback.
> >
> > The error happens in acpi_bus_get_device(), after we read the adr but
> > before we check it, so wondering if we shouldn't revert the checks. Can
> > you try the diff below? I am not sure why there is a crash and we should
> > root-cause this issue, just trying to triangulate what is happening.
> >
> > diff --git a/drivers/soundwire/intel_init.c b/drivers/soundwire/intel_init.c
> > index cabdadb09a1b..6bc87a682fb3 100644
> > --- a/drivers/soundwire/intel_init.c
> > +++ b/drivers/soundwire/intel_init.c
> > @@ -369,13 +369,6 @@ static acpi_status sdw_intel_acpi_cb(acpi_handle
> > handle, u32 level,
> >          if (ACPI_FAILURE(status))
> >                  return AE_OK; /* keep going */
> >
> > -       if (acpi_bus_get_device(handle, &adev)) {
> > -               pr_err("%s: Couldn't find ACPI handle\n", __func__);
> > -               return AE_NOT_FOUND;
> > -       }
> > -
> > -       info->handle = handle;
> > -
> >          /*
> >           * On some Intel platforms, multiple children of the HDAS
> >           * device can be found, but only one of them is the SoundWire
> > @@ -386,6 +379,13 @@ static acpi_status sdw_intel_acpi_cb(acpi_handle
> > handle, u32 level,
> >          if (FIELD_GET(GENMASK(31, 28), adr) != SDW_LINK_TYPE)
> >                  return AE_OK; /* keep going */
> >
> > +       if (acpi_bus_get_device(handle, &adev)) {
> > +               pr_err("%s: Couldn't find ACPI handle\n", __func__);
> > +               return AE_NOT_FOUND;
> > +       }
> > +
> > +       info->handle = handle;
> > +
> >          /* device found, stop namespace walk */
> >          return AE_CTRL_TERMINATE;
> >   }
>
> still the same crash

The modification doesn't fundamentally change the conditions, but
since the flow gets to the acpi_bus_get_device() evaluation, adr
appears to make sense (which is kind of strange, because it shouldn't
in the "no SoundWire" case).