Re: PME via interrupt or SCI mechanism?

["Rafael J. Wysocki" <rjw@sisk.pl>]

On Friday, September 30, 2011, Sarah Sharp wrote:
> On Thu, Sep 29, 2011 at 11:51:49PM +0200, Rafael J. Wysocki wrote:
> > On Thursday, September 29, 2011, Rafael J. Wysocki wrote:
> > > On Thursday, September 29, 2011, Rafael J. Wysocki wrote:
> > > > On Thursday, September 29, 2011, Sarah Sharp wrote:
> > > > > On Thu, Sep 29, 2011 at 09:39:56PM +0200, Rafael J. Wysocki wrote:
> > > > > > On Thursday, September 29, 2011, Sarah Sharp wrote:
> > > > > > > On Thu, Sep 29, 2011 at 12:21:28AM +0200, Rafael J. Wysocki wrote:
> > > > > > Please try the appended patch and check if you see the "Notification error
> > > > > > for GPE" message (please keep your previous debug patches applied).
> > > > > 
> > > > > Do I need to have the ACPI debug_level or debug_layer set to anything in
> > > > > particular to see this message?
> > > > 
> > > > No, I don't think so, but just in case please try the patch below instead
> > > > of the previous one.
> > > 
> > > Actually, please don't, it's a BIOS-related issue after all.  Apparently,
> > > wakeup from xHCD is not supported by the BIOS, because the DSDT defines
> > > the _L0D method for GPE 0D (13), which is the following:
> > > 
> > >         Method (_L0D, 0, NotSerialized)
> > >         {
> > >             Notify (\_SB.PCI0.EHC1, 0x02)
> > >             Notify (\_SB.PCI0.EHC2, 0x02)
> > >             Notify (\_SB.PCI0.HDEF, 0x02)
> > >             Notify (\_SB.PCI0.GLAN, 0x02)
> > >         }
> > > 
> > > so it notifies some devices, but not the xHCD.
> 
> Ok, I'll let our BIOS guys know they need to add the XHC Notify line.

As stated below, it actually may be better to remove the _L0D method
entirely, because our ACPICA GPE handling code should take care of
notifying all of the devices associated with GPE 0D in that case.

> I should see if the x220 ACPI tables have a similar issue or it's a
> different issue with D3.

Yes, that would be good to know.

> > > We might work around this by doing what Matthew has suggested (ie. polling
> > > all PCI and PCIe devices to check if they have PME pending) or perhaps
> > > we can do something about this in ACPICA.
> > > 
> > > Still, the right fix is to put Notify () for the ACPI objects
> > > corresponding to xHCD into the above method.
> > 
> > Or to remove this method altogether (in which case our ACPICA code
> > should take care of the notifications).
> > 
> > Below is a patch (untested!) that should help, but it's kind of
> > suboptimal.  Nevertheless, please try if it works for you.
> 
> The patch seems to finally get the host controller out of D3 when a
> remote wakeup is triggered.

OK

I'll try to prepare a better patch for you to test.

> I still see a lot of log spam from pci_acpi_wake_dev being called for ehci,
> e1000e, and snd_hcda_intel, but eventually the xHCI host gets put in D0.
> dmesg attached.

That's to be expected.  All of the devices mentioned in the _L0D method
above will be notified every time any of devices associated with GPE 0D
sends a PME message to the Root Complex, which happens quite often
while the polling used to check the PME bits is done only every second.

Thanks,
Rafael