Unrecoverable AER error when resuming from RAM (hda regression in 5.7-rc2)

[Takashi Iwai]

On Wed, 22 Apr 2020 22:50:28 +0200,
Bjorn Helgaas wrote:
> 
> [+cc Rafael, linux-pm]
> 
> On Tue, Apr 21, 2020 at 03:08:44PM -0400, Alex Xu (Hello71) wrote:
> > With 5.7-rc2, after resuming from suspend to RAM, I get:
> > 
> > [   55.679382] pcieport 0000:00:03.1: AER: Multiple Uncorrected (Non-Fatal) error received: 0000:00:00.0
> > [   55.679405] pcieport 0000:00:03.1: AER: PCIe Bus Error: severity=Uncorrected (Non-Fatal), type=Transaction Layer, (Requester ID)
> > [   55.679410] pcieport 0000:00:03.1: AER:   device [1022:1453] error status/mask=00100000/04400000
> > [   55.679414] pcieport 0000:00:03.1: AER:    [20] UnsupReq               (First)
> > [   55.679417] pcieport 0000:00:03.1: AER:   TLP Header: 40000004 0a0000ff fffc0e80 00000000
> > [   55.679423] amdgpu 0000:0a:00.0: AER: can't recover (no error_detected callback)
> > [   55.679425] snd_hda_intel 0000:0a:00.1: AER: can't recover (no error_detected callback)
> > [   55.679455] pcieport 0000:00:03.1: AER: device recovery failed
> 
> I'm not at all confident in my decoding skills, but I *think* the TLP
> header decodes to:
> 
>   Fmt           010b         3 DW header with data (32-bit address)
>   Type          00000b       MWr
>   Length        0x4          4 DW = 16 bytes
>   Requester ID  0x0a00       0a:00.0
>   Byte enables  0xff
>   Address       0xfffc0e80
> 
> which would mean the 0a:00.0 GPU did a 16-byte write to 0xfffc0e80,
> and the 00:03.1 Root Port reported that as an Unsupported Request.
> I don't know why that would be unless the address is invalid.
> 
> Maybe that's supposed to be an MSI address?  Maybe a complete dmesg or
> /proc/iomem would have a clue?
> 
> I feel like this UR issue could be a PCI core issue or maybe some sort
> of misuse of PCI power management, but I can't seem to get traction on
> it.
> 
> > Then the display freezes and the system basically falls apart (can't 
> > even sudo reboot -f, need to use magic sysrq).
> > 
> > I bisected this to "ALSA: hda: Skip controller resume if not needed". 
> > Setting snd_hda_intel.power_save=0 resolves the issue.
> 
> FWIW, the complete citation is c4c8dd6ef807 ("ALSA: hda: Skip
> controller resume if not needed"),
> https://git.kernel.org/linus/c4c8dd6ef807, which first appeared in
> v5.7-rc2.

Yes, and I posted the fix patch right now:
  https://lore.kernel.org/r/20200422203744.26299-1-tiwai@suse.de

The possible cause was the tricky resume code that both HD-audio
controller (the parent PCI device) and the codec devices used.

At least the patch above seems working for the reporter's machine.
Now we need a bit more testing before merging, but it looks promising,
so far.


thanks,

Takashi