Re: PCI: Revert "PCI: Add runtime PM support for PCIe ports"

[Peter Wu <peter@lekensteyn.nl>]

On Wed, Jan 04, 2017 at 09:16:39AM +0100, Lukas Wunner wrote:
> On Tue, Jan 03, 2017 at 06:05:57PM -0600, Bjorn Helgaas wrote:
> > I don't *want* to apply the revert.  It's on my for-linus branch as a
> > worst-case scenario change if we can't figure out a better fix.
> > 
> > The patch below is preferable, but I'd rather not take even it,
> > because it takes away functionality and forces people to use a boot
> > parameter to restore it.  I expect that somebody will figure out how
> > to fix the regression Kilian found and also keep the new functionality
> > (without requiring boot parameters) before v4.10.
> 
> The issue is constrained to hybrid graphics laptops with Nvidia discrete
> GPU using nouveau.  Hence it needs to be fixed in nouveau, not in the
> PCI core.

The problem is not necessarily in the nouveau driver, the same problem
occurs when you enable RPM without loading nouveau. The issue is limited
though to some newer hybrid graphics laptops with Nvidia GPUs. While a
quirk can be added to nouveau, I think that a (temporary) quirk in core
would also be reasonable (since it also occurs without nouveau).

> (AFAIUI, laptops with AMD discrete GPU are not affected as it is known
> when and how to call an ACPI method versus using PR3.)
> 
> (Neither are laptops using the Nvidia proprietary driver as it doesn't
> runtime suspend the card.  But battery life will be terrible then.)
> 
> We're at rc2 so the time frame for coming up with a fix is probably
> 4 weeks.  Peter and others have tried for months to reverse-engineer
> how to handle runtime PM on newer Nvidia cards.  It seems likely that
> we'll not find the ultimate solution to the problem within 4 weeks.

Yep, a quick proper fix seems unlikely.
[ Help/ideas are welcome, I suspect that these failures to restore power
on laptops designed for Win8+ all have the same cause, related to some
unknown interaction between ACPI and PCI. Some links:
https://bugzilla.kernel.org/show_bug.cgi?id=190861
https://bugzilla.kernel.org/show_bug.cgi?id=156341 ]

> The way it is now, i.e. defaulting to PR3 when available, regresses
> certain laptops such as Kilian's.  If on the other hand we default to
> DSM when available, we'll regress certain other laptops, as Peter has
> pointed out.  Whitelisting or blacklisting laptops doesn't seem a good
> approach either, ideally we'd want to use PR3 as Windows does.
> 
> As said, the only short-term solution I see is to add an "optimus"
> module_param to nouveau to allow users to select which method to use.
> So in Kilian's case an additional command line parameter would be
> necessary to fix the issue.
> 
> Does anyone see a better solution or can we agree on this one?  If so
> I can come up with a patch.  This could go in via Dave Airlie's tree.

As pcie_port_pm=off already reverts to DSM, I do not think that an
additional (temporary) nouveau module parameter is going to help. I
instead propose a (hopefully temporary) quirk in pci core that disables
D3cold RPM for just Kilians Lenovo laptop (basically defaulting to
pcie_port_pm=off). Then the option pcie_port_pm=force can still be used
to test possible solutions in the future.
-- 
Kind regards,
Peter Wu
https://lekensteyn.nl