On Wed, 18 Apr 2012, Steven Rostedt wrote:
> On Wed, 2012-04-18 at 23:10 +0600, Andrey Rahmatullin wrote:
> > On Wed, Apr 18, 2012 at 12:41:10PM -0400, Alan Stern wrote:
> > > Here's the situation. When the script unbinds ehci-hcd,
> > > pci_device_remove() sets the current state to PCI_UNKNOWN. The patch
> > > to hcd-pci.c does the same thing before the suspend_noirq phase begins,
> > > i.e., before pci_prepare_to_sleep() is called.
> > >
> > > In pci_raw_set_power_state(), there's a "switch" statement that depends
> > > on dev->current_state. If the current state is PCI_D0 then the new
> > > state in pmcsr is set to PCI_D3; if the current state is PCI_UNKNOWN
> > > then the state in pmcsr is left unchanged. After that, the value in
> > > pmcsr is written to the controller.
> > >
> > > This means that when ehci-hcd is unbound or the patch is installed, the
> > > controller stays in D0. That's why we see those "Refused to change
> > > power state" messages, since D0 is not what the target state was
> > > supposed to be. When ehci-hcd remains bound and the patch isn't
> > > installed, the controller is put into D3.
> > >
> > > And then finally, the computer crashes during the final stage of
> > > suspend if either controller is in D3. Clearly this is a bug in the
> > > firmware, not in the kernel.
> Is there a way to detect this chipset or something, to add a workaround
> for it?
Yes, there are ways to work around this. At the moment I'm not sure
what would be best; we can ask Rafael. One big remaining puzzle: Why
are the EHCI controllers the only devices that cause a crash when left
in D3 during suspend? We may never know...
In the meantime, just to be certain of the diagnosis, here's a
different patch for you guys to try. This will test the ehci-hcd
unbound path (i.e., use it with the script). The patch removes the
line that sets the dev->current_state to PCI_UNKNOWN when the driver is
unbound. Thus current_state will remain equal to PCI_D0, so
pci_prepare_to_sleep should put the controllers into D3, which we
expect will cause a crash.
Please try this both with and without pm_test set to "platform", and
post the debugging dmesg output from whichever cases the computer
survives.
Alan Stern
Index: usb-3.4/drivers/pci/pci-driver.c
===================================================================
--- usb-3.4.orig/drivers/pci/pci-driver.c
+++ usb-3.4/drivers/pci/pci-driver.c
@@ -394,8 +394,8 @@ static int pci_device_remove(struct devi
* If the device is still on, set the power state as "unknown",
* since it might change by the next time we load the driver.
*/
- if (pci_dev->current_state == PCI_D0)
- pci_dev->current_state = PCI_UNKNOWN;
+// if (pci_dev->current_state == PCI_D0)
+// pci_dev->current_state = PCI_UNKNOWN;
/*
* We would love to complain here if pci_dev->is_enabled is set, that
@@ -713,6 +713,8 @@ static int pci_pm_suspend_noirq(struct d
if (!pm) {
pci_save_state(pci_dev);
+ pci_prepare_to_sleep(pci_dev);
+ pci_pm_set_unknown_state(pci_dev);
return 0;
}
Index: usb-3.4/drivers/pci/pci.c
===================================================================
--- usb-3.4.orig/drivers/pci/pci.c
+++ usb-3.4/drivers/pci/pci.c
@@ -1710,6 +1710,7 @@ pci_power_t pci_target_state(struct pci_
*/
int pci_prepare_to_sleep(struct pci_dev *dev)
{
+ pci_power_t cur_state = dev->current_state;
pci_power_t target_state = pci_target_state(dev);
int error;
@@ -1723,6 +1724,8 @@ int pci_prepare_to_sleep(struct pci_dev
if (error)
pci_enable_wake(dev, target_state, false);
+ dev_info(&dev->dev, "cur %d target %d error %d\n", cur_state,
+ target_state, error);
return error;
}
_______________________________________________
linux-pm mailing list
linux-pm@xxxxxxxxxxxxxxxxxxxxxxxxxx
https://lists.linuxfoundation.org/mailman/listinfo/linux-pm
