Re: [PATCH v2] PCI/PM: Avoid suspending the device with errors

[Raag Jadav <raag.jadav@intel.com>]

On Thu, Apr 24, 2025 at 01:02:58PM +0200, Rafael J. Wysocki wrote:
> On Thu, Apr 24, 2025 at 7:38 AM Raag Jadav <raag.jadav@intel.com> wrote:
> >
> > On Wed, Apr 23, 2025 at 02:41:52PM +0200, Rafael J. Wysocki wrote:
> > > On Tue, Apr 22, 2025 at 3:55 PM Raag Jadav <raag.jadav@intel.com> wrote:
> > > >
> > > > If an error is triggered before or during system suspend, any bus level
> > > > power state transition will result in unpredictable behaviour by the
> > > > device with failed recovery. Avoid suspending such a device and leave
> > > > it in its existing power state.
> > > >
> > > > This only covers the devices that rely on PCI core PM for their power
> > > > state transition.
> > > >
> > > > Signed-off-by: Raag Jadav <raag.jadav@intel.com>
> > > > ---
> > > >
> > > > v2: Synchronize AER handling with PCI PM (Rafael)
> > > >
> > > > More discussion on [1].
> > > > [1] https://lore.kernel.org/all/CAJZ5v0g-aJXfVH+Uc=9eRPuW08t-6PwzdyMXsC6FZRKYJtY03Q@mail.gmail.com/
> > > >
> > > >  drivers/pci/pci-driver.c |  3 ++-
> > > >  drivers/pci/pcie/aer.c   | 11 +++++++++++
> > > >  include/linux/aer.h      |  2 ++
> > > >  3 files changed, 15 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/drivers/pci/pci-driver.c b/drivers/pci/pci-driver.c
> > > > index f57ea36d125d..289a1fa7cb2d 100644
> > > > --- a/drivers/pci/pci-driver.c
> > > > +++ b/drivers/pci/pci-driver.c
> > > > @@ -884,7 +884,8 @@ static int pci_pm_suspend_noirq(struct device *dev)
> > > >                 }
> > > >         }
> > > >
> > > > -       if (!pci_dev->state_saved) {
> > > > +       /* Avoid suspending the device with errors */
> > > > +       if (!pci_aer_in_progress(pci_dev) && !pci_dev->state_saved) {
> > >
> > > Apart from the potential raciness mentioned by Bjorn, doing it just
> > > here is questionable because this is not the only place where the PCI
> > > device power state can change.
> > >
> > > It would be better to catch this in pci_set_low_power_state() IMO.
> >
> > I'm not sure if we should prevent power state transition for the users
> > that explicitly want to transition.
> >
> > Also, the device state can potentially be corrupted because of the errors,
> > so we'd probably want to avoid pci_save_state() as well, which is what
> > I attempted here.
> 
> But it's not what the changelog is saying.
> 
> If you want to avoid pci_save_state(), there are also other places
> when it is called and then you also may want to avoid the transition
> because if the state is not saved, it won't be possible to restore it.

Yes, above logic will skip both pci_save_state() and pci_prepare_to_sleep()
and in turn set skip_bus_pm flag to prevent power state transition.

Considering we're targeting only system PM cases, I could find pci_save_state()
being used in pci_pm_suspend_noirq() and pci_pm_freeze_noirq() (excluding
legacy suspend). pci_pm_freeze_noirq() doesn't seem to attempt any power state
transition by itself, so any other cases that you think are worth being covered
here?

Raag