Re: Discrete Nvidia GPU shutdown with recent kernel / BIOS (on T440s laptop) - Should be automatic with runtime pm now ?

On Wed, 22 Jul 2015, Sylvain Munaut wrote:

Hi Alan,


Thanks for the details.

quoted

quoted

tl;dr : Should device allowed to do runtime pm that have no driver go
to D3 on their own ?

Quick answer: See the comments at the start of local_pci_probe() in
drivers/pci/pci-driver.c.  PCI devices without a kernel driver are
always left in D0, regardless of the runtime PM status.  This is
because userspace may try to use the device, and the kernel wouldn't
know to take the device out of D3 when that happens.

Does linux expose some way for userspace to transition the device to
another power state without the need for a driver ?

I don't think so.  Of course, such a thing could be added.  One simple 
way to do it would be to create a "do-nothing" driver that would allow 
userspace to access the device directly and would include runtime PM 
suport.

I looked at the docs and tried poking some of the sysfs files, but
except enabling/disabling the runtme pm, I couldn't really change
anything, but maybe I missed something.

You probably didn't miss anything.

quoted

quoted

 And should pci root bridge go to D3 when all the
device on them are on D3 ?

I don't know the answer to that.  It's probably closely related to the
issue of whether PCI bridges need to have a kernel driver.

OK.

Do you know if there could be any adverse effect of transitioning it myself ?

Basically I'm doing this when turning off the card :

struct pci_dev *port_dev = pci_upstream_bridge(pdev);

pci_save_state(port_dev);
pci_clear_master(port_dev);
pci_disable_device(port_dev);
pci_set_power_state(port_dev, PCI_D3hot);


And this when waking it up :

pci_set_power_state(port_dev, PCI_D0);
pci_restore_state(port_dev);
pci_enable_device(port_dev);
pci_set_master(port_dev);

(I removed the return value checking to make the code more concise here).

That seems reasonable.  But I am far from an expert on PCI or Linux's
PCI subsystem.  You really should address this sort of question to the
linux-pci mailing list.

Alan Stern