Re: [PATCH v4] pci: prevent putting nvidia GPUs into lower device states on certain intel bridges

[Karol Herbst]

On Wed, Nov 20, 2019 at 12:48 PM Rafael J. Wysocki <rafael@xxxxxxxxxx> wrote:
>
> On Wed, Nov 20, 2019 at 12:22 PM Mika Westerberg
> <mika.westerberg@xxxxxxxxx> wrote:
> >
> > On Wed, Nov 20, 2019 at 11:52:22AM +0100, Rafael J. Wysocki wrote:
> > > On Wed, Nov 20, 2019 at 11:18 AM Mika Westerberg
> > > <mika.westerberg@xxxxxxxxx> wrote:
> > > >
> > > > Hi Karol,
> > > >
> > > > On Tue, Nov 19, 2019 at 11:26:45PM +0100, Karol Herbst wrote:
> > > > > On Tue, Nov 19, 2019 at 10:50 PM Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote:
> > > > > >
> > > > > > [+cc Dave]
> > > > > >
> > > > > > On Thu, Oct 17, 2019 at 02:19:01PM +0200, Karol Herbst wrote:
> > > > > > > Fixes state transitions of Nvidia Pascal GPUs from D3cold into higher device
> > > > > > > states.
> > > > > > >
> > > > > > > v2: convert to pci_dev quirk
> > > > > > >     put a proper technical explanation of the issue as a in-code comment
> > > > > > > v3: disable it only for certain combinations of intel and nvidia hardware
> > > > > > > v4: simplify quirk by setting flag on the GPU itself
> > > > > >
> > > > > > I have zero confidence that we understand the real problem, but we do
> > > > > > need to do something with this.  I'll merge it for v5.5 if we get the
> > > > > > minor procedural stuff below straightened out.
> > > > > >
> > > > >
> > > > > Thanks, and I agree with your statement, but at this point I think
> > > > > only Intel can help out digging deeper as I see no way to debug this
> > > > > further.
> > > >
> > > > I don't have anything against this patch, as long as the quirk stays
> > > > limited to the particular root port leading to the NVIDIA GPU. The
> > > > reason why I think it should to be limited is that I'm pretty certain
> > > > the problem is not in the root port itself. I have here a KBL based
> > > > Thinkpad X1 Carbon 6th gen that can put the TBT controller into D3cold
> > > > (it is connected to PCH root port) and it wakes up there just fine, so
> > > > don't want to break that.
> > > >
> > > > Now, PCIe devices cannot go into D3cold all by themselves. They always
> > > > need help from the platform side which is ACPI in this case. This is
> > > > done by having the device to have _PR3 method that returns one or more
> > > > power resources that the OS is supposed to turn off when the device is
> > > > put into D3cold. All of that is implemented as form of ACPI methods that
> > > > pretty much do the hardware specific things that are outside of PCIe
> > > > spec to get the device into D3cold. At high level the _OFF() method
> > > > causes the root port to broadcast PME_Turn_Off message that results the
> > > > link to enter L2/3 ready, it then asserts PERST, configures WAKE (both
> > > > can be GPIOs) and finally removes power (if the link goes into L3,
> > > > otherwise it goes into L2).
> > > >
> > > > I think this is where the problem actually lies - the ASL methods that
> > > > are used to put the device into D3cold and back. We know that in Windows
> > > > this all works fine so unless Windows quirks the root port the same way
> > > > there is another reason behind this.
> > > >
> > > > In case of Dell XPS 9560 (IIRC that's the machine you have) the
> > > > corresponding power resource is called \_SB.PCI0.PEG0.PG00 and its
> > > > _ON/_OFF methods end up calling PGON()/PGOF() accordingly. The methods
> > > > itself do lots of things and it is hard to follow the dissassembled
> > > > ASL which does not have any comments but there are couple of things that
> > > > stand out where we may go into a different path. One of them is this in
> > > > the PGOF() method:
> > > >
> > > >    If (((OSYS <= 0x07D9) || ((OSYS == 0x07DF) && (_REV == 0x05))))
> > > >
> > > > The ((OSYS == 0x07DF) && (_REV == 0x05)) checks specifically for Linux
> > > > (see [1] and 18d78b64fddc ("ACPI / init: Make it possible to override
> > > > _REV")) so it might be that Dell people tested this at some point in
> > > > Linux as well. Added Mario in case he has any ideas.
> > > >
> > > > Previously I suggested you to try the ACPI method tracing to see what
> > > > happens inside PGOF(). Did you have time to try it? It may provide more
> > > > information about that is happening inside those methods and hopefully
> > > > point us to the root cause.
> > > >
> > > > Also if you haven't tried already passing acpi_rev_override in the
> > > > command line makes the _REV to return 5 so it should go into the "Linux"
> > > > path in PGOF().
> > >
> > > Oh, so does it look like we are trying to work around AML that tried
> > > to work around some problematic behavior in Linux at one point?
> >
> > Yes, it looks like so if I read the ASL right.
>
> OK, so that would call for a DMI-based quirk as the real cause for the
> issue seems to be the AML in question, which means a firmware problem.
>

And I disagree as this is a linux specific workaround and windows goes
that path and succeeds. This firmware based workaround was added,
because it broke on Linux.