On Fri, 2020-08-21 at 12:32 +0300, Mika Westerberg wrote:
> Hi,
>
> On Thu, Aug 20, 2020 at 11:36:37AM -0400, Lyude Paul wrote:
> > On Thu, 2020-08-20 at 10:13 +0200, Lukas Wunner wrote:
> > > On Wed, Aug 19, 2020 at 04:06:25PM +0300, Mika Westerberg wrote:
> > > > Sec 7.5.3.6 requires such Ports to support DLL Link Active reporting,
> > > > but
> > > > at least the Intel JHL6240 Thunderbolt 3 Bridge [8086:15c0] and the
> > > > Intel
> > > > JHL7540 Thunderbolt 3 Bridge [8086:15ea] do not.
> > > [...]
> > > > + * Also do the same for devices that have power management
> > > > disabled
> > > > + * by their driver and are completely power managed through the
> > > > + * root port power resource instead. This is a special case for
> > > > + * nouveau.
> > > > */
> > > > - if (!pci_is_pcie(dev)) {
> > > > + if (!pci_is_pcie(dev) || !child->pm_cap) {
> > >
> > > It sounds like the above-mentioned Thunderbolt controllers are broken,
> > > not the Nvidia cards, so to me (as an outside observer) it would seem
> > > more logical that a quirk for the former is needed. The code comment
> > > suggests that nouveau somehow has a problem, but that doesn't seem to
> > > be the case (IIUC). Also, it's a little ugly to have references to
> > > specific drivers in PCI core code.
> > >
> > > Maybe this can be fixed with quirks for the Thunderbolt controllers
> > > which set a flag, and that flag causes the 1000 msec wait to be skipped?
> >
> > Sorry, some stuff came up yesterday so I didn't get the time to go through
> > my
> > laptops and test them. I do agree with this though - I'd be worried as well
> > that
> > nouveau might not be the only driver out there that needs this kind of delay
>
> I actually expect that nouveau is the only one because it is doing some
> PM tricks to get the runtime PM working, which is that it leaves the GPU
> device in D0 and puts the parent root port into D3cold. The BIOS ASL
> code has some assumptions there and I think this 1000 ms delay just
> works that around by luck ;-)
>
> IIRC Bjorn suggested quirking the affected downstream ports when I
> originally sent the patch but I thought we could make this solution more
> generic. Which of course, did not work too well.
>
> I can look into the quirk solution instead if this is what people
> prefer.
yeah, probably the safest bet IMO.
>
--
Sincerely,
Lyude Paul (she/her)
Software Engineer at Red Hat
