Re: [PATCH v3] PCI/PM: Only read PCI_PM_CTRL register when available

[Bjorn Helgaas <helgaas@kernel.org>]

On Fri, Aug 25, 2023 at 11:57:00AM +0800, Feiyang Chen wrote:
> On Fri, Aug 25, 2023 at 5:59 AM Bjorn Helgaas <helgaas@kernel.org> wrote:
> > On Thu, Aug 24, 2023 at 09:37:38AM +0800, Feiyang Chen wrote:
> > > When the current state is already PCI_D0, pci_power_up() will return
> > > 0 even though dev->pm_cap is not set. In that case, we should not
> > > read the PCI_PM_CTRL register in pci_set_full_power_state().
> > >
> > > There is nothing more needs to be done below in that case.
> > > Additionally, pci_power_up() has two callers only and the other one
> > > ignores the return value, so we can safely move the current state
> > > check from pci_power_up() to pci_set_full_power_state().
> >
> > Does this fix a bug?  I guess it does, because previously
> > pci_set_full_power_state() did a config read at 0 + PCI_PM_CTRL, i.e.,
> > offset 4, which is actually PCI_COMMAND, and set dev->current_state
> > based on that.  So dev->current_state is now junk, right?
> 
> Yes.
> 
> > This might account for some "Refused to change power state from %s to D0"
> > messages.
> >
> > How did you find this?  It's nice if we can mention a symptom so
> > people can connect the problem with this fix.
> 
> We are attempting to add MSI support for our stmmac driver, but the
> pci_alloc_irq_vectors() function always fails.
> After looking into it more, we came across the message "Refused to
> change power state from D3hot to D0" :)

So I guess this device doesn't have a PM Capability at all?  Can you
collect the "sudo lspci -vv" output?  The PM Capability is required
for all PCIe devices, so maybe this is a conventional PCI device?

> > This sounds like something that probably should have a stable tag?
> 
> Do I need to include the symptom and Cc in the commit message and
> then send v4?

> > > Fixes: e200904b275c ("PCI/PM: Split pci_power_up()")
> > > Signed-off-by: Feiyang Chen <chenfeiyang@loongson.cn>
> > > Reviewed-by: Rafael J. Wysocki <rafael@kernel.org>
> > > ---
> > >  drivers/pci/pci.c | 9 +++++----
> > >  1 file changed, 5 insertions(+), 4 deletions(-)
> > >
> > > diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
> > > index 60230da957e0..7e90ab7b47a1 100644
> > > --- a/drivers/pci/pci.c
> > > +++ b/drivers/pci/pci.c
> > > @@ -1242,9 +1242,6 @@ int pci_power_up(struct pci_dev *dev)
> > >               else
> > >                       dev->current_state = state;
> > >
> > > -             if (state == PCI_D0)
> > > -                     return 0;
> > > -
> > >               return -EIO;
> > >       }
> > >
> > > @@ -1302,8 +1299,12 @@ static int pci_set_full_power_state(struct pci_dev *dev)
> > >       int ret;
> > >
> > >       ret = pci_power_up(dev);
> > > -     if (ret < 0)
> > > +     if (ret < 0) {
> > > +             if (dev->current_state == PCI_D0)
> > > +                     return 0;
> > > +
> > >               return ret;
> > > +     }
> > >       pci_read_config_word(dev, dev->pm_cap + PCI_PM_CTRL, &pmcsr);
> > >       dev->current_state = pmcsr & PCI_PM_CTRL_STATE_MASK;

One thing that makes me hesitate a little bit is that we rely on the
failure return from pci_power_up() to guard the dev->pm_cap usage.
That's slightly obscure, and I liked the way the v1 patch made it
explicit.

And it seems slightly weird that when there's no PM cap,
pci_power_up() always returns failure even if the platform was able to
put the device in D0.

Anyway, here's a proposal for commit log and updated comment for
pci_power_up():


commit 5694ba13b004 ("PCI/PM: Only read PCI_PM_CTRL register when available")
Author: Feiyang Chen <chenfeiyang@loongson.cn>
Date:   Thu Aug 24 09:37:38 2023 +0800

    PCI/PM: Only read PCI_PM_CTRL register when available
    
    For a device with no Power Management Capability, pci_power_up() previously
    returned 0 (success) if the platform was able to put the device in D0,
    which led to pci_set_full_power_state() trying to read PCI_PM_CTRL, even
    though it doesn't exist.
    
    Since dev->pm_cap == 0 in this case, pci_set_full_power_state() actually
    read the wrong register, interpreted it as PCI_PM_CTRL, and corrupted
    dev->current_state.  This led to messages like this in some cases:
    
      pci 0000:01:00.0: Refused to change power state from D3hot to D0
    
    To prevent this, make pci_power_up() always return a negative failure code
    if the device lacks a Power Management Capability, even if non-PCI platform
    power management has been able to put the device in D0.  The failure will
    prevent pci_set_full_power_state() from trying to access PCI_PM_CTRL.
    
    Fixes: e200904b275c ("PCI/PM: Split pci_power_up()")
    Link: https://lore.kernel.org/r/20230824013738.1894965-1-chenfeiyang@loongson.cn
    Signed-off-by: Feiyang Chen <chenfeiyang@loongson.cn>
    Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
    Reviewed-by: "Rafael J. Wysocki" <rafael@kernel.org>
    Cc: stable@vger.kernel.org	# v5.19+


diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
index 60230da957e0..39728196e295 100644
--- a/drivers/pci/pci.c
+++ b/drivers/pci/pci.c
@@ -1226,6 +1226,10 @@ static int pci_dev_wait(struct pci_dev *dev, char *reset_type, int timeout)
  *
  * On success, return 0 or 1, depending on whether or not it is necessary to
  * restore the device's BARs subsequently (1 is returned in that case).
+ *
+ * On failure, return a negative error code.  Always return failure if @dev
+ * lacks a Power Management Capability, even if the platform was able to
+ * put the device in D0 via non-PCI means.
  */
 int pci_power_up(struct pci_dev *dev)
 {
@@ -1242,9 +1246,6 @@ int pci_power_up(struct pci_dev *dev)
 		else
 			dev->current_state = state;
 
-		if (state == PCI_D0)
-			return 0;
-
 		return -EIO;
 	}
 
@@ -1302,8 +1303,12 @@ static int pci_set_full_power_state(struct pci_dev *dev)
 	int ret;
 
 	ret = pci_power_up(dev);
-	if (ret < 0)
+	if (ret < 0) {
+		if (dev->current_state == PCI_D0)
+			return 0;
+
 		return ret;
+	}
 
 	pci_read_config_word(dev, dev->pm_cap + PCI_PM_CTRL, &pmcsr);
 	dev->current_state = pmcsr & PCI_PM_CTRL_STATE_MASK;