amd64-agp: resume but too late?

amd64-agp: resume but too late?

[Pavel Machek]

Hi!

I don't get it:

--- a/drivers/char/agp/amd64-agp.c
+++ b/drivers/char/agp/amd64-agp.c
@@ -549,6 +549,8 @@ static int agp_amd64_suspend(struct pci_
        return 0;
 }
 
+/* FIXME: this is strange; we enable iommu translation here, but that
+ * may be too late, no? */
 static int agp_amd64_resume(struct pci_dev *pdev)
 {
        pci_set_power_state(pdev, PCI_D0);

...but resume is normal pci device, so other devices may be resumed
before this one, and already depend on iommu running...

How is this supposed to work? Should agp_amd64_resume be called from
sysdev or something like that?
								Pavel
-- 
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html

Re: amd64-agp: resume but too late?

[Thomas Gleixner]

On Mon, 26 May 2008, Pavel Machek wrote:
> Hi!
> 
> I don't get it:
> 
> --- a/drivers/char/agp/amd64-agp.c
> +++ b/drivers/char/agp/amd64-agp.c
> @@ -549,6 +549,8 @@ static int agp_amd64_suspend(struct pci_
>         return 0;
>  }
>  
> +/* FIXME: this is strange; we enable iommu translation here, but that
> + * may be too late, no? */
>  static int agp_amd64_resume(struct pci_dev *pdev)
>  {
>         pci_set_power_state(pdev, PCI_D0);
> 
> ...but resume is normal pci device, so other devices may be resumed
> before this one, and already depend on iommu running...
> 
> How is this supposed to work? Should agp_amd64_resume be called from
> sysdev or something like that?

Hmm, you sent a patch some days ago to change it to
suspend_late/resume_early. Is the discussion resolved about that ?

Thanks,
	tglx

Re: amd64-agp: resume but too late?

[Pavel Machek]

On Tue 2008-05-27 14:18:19, Thomas Gleixner wrote:
> On Mon, 26 May 2008, Pavel Machek wrote:
> > Hi!
> > 
> > I don't get it:
> > 
> > --- a/drivers/char/agp/amd64-agp.c
> > +++ b/drivers/char/agp/amd64-agp.c
> > @@ -549,6 +549,8 @@ static int agp_amd64_suspend(struct pci_
> >         return 0;
> >  }
> >  
> > +/* FIXME: this is strange; we enable iommu translation here, but that
> > + * may be too late, no? */
> >  static int agp_amd64_resume(struct pci_dev *pdev)
> >  {
> >         pci_set_power_state(pdev, PCI_D0);
> > 
> > ...but resume is normal pci device, so other devices may be resumed
> > before this one, and already depend on iommu running...
> > 
> > How is this supposed to work? Should agp_amd64_resume be called from
> > sysdev or something like that?
> 
> Hmm, you sent a patch some days ago to change it to
> suspend_late/resume_early. Is the discussion resolved about that ?

I'm not aware of such patch. I don't think I sent it.

I did add suspend/resume support to pci-gart_64.c , which is
confusingly similar code. That patch needs a bit more testing, and
some de-uglyfication.
									Pavel
-- 
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html