On Tue, Aug 06, 2024 at 06:11:07PM +0530, Manivannan Sadhasivam wrote:
> On Tue, Aug 06, 2024 at 08:53:01AM +0200, Lukas Wunner wrote:
> > AFAICS we always program the device to go to D3hot and the platform
> > then cuts power, thereby putting it into D3cold. So D3hot is never
> > skipped. See __pci_set_power_state():
> >
> > if (state == PCI_D3cold) {
> > /*
> > * To put the device in D3cold, put it into D3hot in the native
> > * way, then put it into D3cold using platform ops.
> > */
> > error = pci_set_low_power_state(dev, PCI_D3hot, locked);
> >
> > if (pci_platform_power_transition(dev, PCI_D3cold))
> > return error;
> >
>
> This is applicable only to pci_set_power_state(), but AFAIK PCIe spec
> doesn't mandate switching to D3Hot for entering D3Cold.
Per PCI Bus Power Management Interface Specification r1.2 sec 5.5 fig 5-1,
the only supported state transition to D3cold is from D3hot.
Per PCIe r6.2 sec 5.2, "PM is compatible with the PCI Bus Power Management
Interface Specification".
Granted, PCI-PM is an ancient spec, so I think anyone can be forgiven
for not knowing its intricacies off-the-cuff. :)
> So the PCIe host controller drivers (especically non-ACPI platforms)
> may just send PME_Turn_Off followed by removing the slot power
> (which again is not controlled by pci_set_power_state())
> as there are no non-ACPI related hooks as of now.
Ideally, devicetree-based platforms should be brought into the
platform_pci_*() fold to align them with ACPI and get common
behavior across all platforms.
Thanks,
Lukas
