Re: [PATCH] PCI: increase D3 delay for AMD Ryzen5/7 XHCI controllers

["Rafael J. Wysocki" <rafael@kernel.org>]

On Tue, Oct 15, 2019 at 7:31 AM Daniel Drake <drake@endlessm.com> wrote:
>
> On Mon, Oct 14, 2019 at 11:43 PM Bjorn Helgaas <helgaas@kernel.org> wrote:
> > Can you tell if this is because the Ryzen7 XHCI controller is out of
> > spec, or is the Linux PCI core missing some delay?  If the latter,
> > fixing the core might fix other devices as well.
> >
> > Mika has this patch:
> > https://lore.kernel.org/r/20190821124519.71594-1-mika.westerberg@linux.intel.com
> > for similar issues, but I think that patch fixes D3cold->D0
> > transitions, and your patch appears to be concerned with D3hot->D0
> > transitions.
>
> It's actually coming out of D3cold here, however what happens right
> before this is that __pci_start_power_transition() calls
> pci_platform_power_transition(D0) to leave D3cold state, then
> pci_update_current_state() reads PMCSR and updates dev->current_state
> to D3hot.

Which pci_update_current_state() do you mean, exactly?

Note that pci_platform_power_transition() itself contains one, which
triggers after a successful change of the ACPI power state of the
device (which should be the case here).

Then, there is one in pci_power_up(), but before that the device's
PMCSR is read from and written to in pci_raw_set_power_state().

It looks like the reads from it after the ACPI power state change are
all successful, but the write isn't unless there is the delay between
it and the former platform_pci_set_power_state().

> The 20ms delay for these XHCI controllers is needed precisely at this
> point - after writing PMCSR to move to D0, and before reading it back
> to check the result.
> I tried moving the delay immediately before writing PMCSR, but that
> doesn't work. Based on that, it seems like it's just a little out of
> spec.

That I agree with and the platform firmware doesn't compensate for
that (which it could do, arguably).