| Message ID | 3623886.MHq7AAxBmi@kreacher (mailing list archive) |
|---|---|
| State | Superseded |
| Delegated to: | Bjorn Helgaas |
| Headers | show |
| Series | : PCI: PM: Improvements related to D3cold -> D0 transitions | expand |
On Mon, Apr 04, 2022 at 05:41:13PM +0200, Rafael J. Wysocki wrote: > From: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > > In theory, pci_power_up() may leave a device in D0-uninitialized > during a transition from D3cold to D0. > > Say, a PCIe device depending on some ACPI power resources is put into > D3cold, so the power resources in question are all turned off. Then, > pci_power_up() is called to put it into D0. > > It first calls pci_platform_power_transition() which invokes > platform_pci_set_power_state() to turn on the ACPI power resources > depended on by the device and, if that is successful, it calls > pci_update_current_state() to update the current_state field of > the PCI device object. If the device's configuration space is > accessible at this point, which is the case if > platform_pci_set_power_state() leaves it in D0-uninitialized (and > there's nothing to prevent it from doing so), current_state will be > set to PCI_D0 and the pci_raw_set_power_state() called subsequently > will notice that the device is in D0 already and do nothing. > However, that is not correct, because it may be still necessary to > restore the device's BARs at this point. > > To address this issue, set current_state temporarily to PCI_D3hot > in the cases in which pci_raw_set_power_state() may need to do more > than just changing the power state of the device. > > Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Reviewed-by: Mika Westerberg <mika.westerberg@linux.intel.com>
On Tue, Apr 5, 2022 at 1:45 PM Mika Westerberg <mika.westerberg@linux.intel.com> wrote: > > On Mon, Apr 04, 2022 at 05:41:13PM +0200, Rafael J. Wysocki wrote: > > From: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > > > > In theory, pci_power_up() may leave a device in D0-uninitialized > > during a transition from D3cold to D0. > > > > Say, a PCIe device depending on some ACPI power resources is put into > > D3cold, so the power resources in question are all turned off. Then, > > pci_power_up() is called to put it into D0. > > > > It first calls pci_platform_power_transition() which invokes > > platform_pci_set_power_state() to turn on the ACPI power resources > > depended on by the device and, if that is successful, it calls > > pci_update_current_state() to update the current_state field of > > the PCI device object. If the device's configuration space is > > accessible at this point, which is the case if > > platform_pci_set_power_state() leaves it in D0-uninitialized (and > > there's nothing to prevent it from doing so), current_state will be > > set to PCI_D0 and the pci_raw_set_power_state() called subsequently > > will notice that the device is in D0 already and do nothing. > > However, that is not correct, because it may be still necessary to > > restore the device's BARs at this point. > > > > To address this issue, set current_state temporarily to PCI_D3hot > > in the cases in which pci_raw_set_power_state() may need to do more > > than just changing the power state of the device. > > > > Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > > Reviewed-by: Mika Westerberg <mika.westerberg@linux.intel.com> Thanks, but on second thought, I'm not sure if this is the best way to address the issue. Basically, pci_power_up() is called in two places, in pci_set_power_state() (for the transitions to D0) and in pci_pm_default_resume_early(). In the latter case, pci_restore_state() is called right after it and that covers BARs restoration, so nothing more needs to be done in that case. This means that pci_set_power_state() is the only place needing to restore the BARs when going into D0 from D3hot or deeper and it is better to move BARs restoration directly into it. I'll update the series accordingly and resend. I also think that the mandatory delay is not needed at all when pci_raw_set_power_state() is called for transitions D3cold -> D0, because in that case either the device has been powered up via platform_pci_set_power_state(), or via the bridge resume which takes the delay into account.
Hi Rafael, On Thu, Apr 07, 2022 at 09:01:59PM +0200, Rafael J. Wysocki wrote: > On Tue, Apr 5, 2022 at 1:45 PM Mika Westerberg > <mika.westerberg@linux.intel.com> wrote: > > > > On Mon, Apr 04, 2022 at 05:41:13PM +0200, Rafael J. Wysocki wrote: > > > From: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > > > > > > In theory, pci_power_up() may leave a device in D0-uninitialized > > > during a transition from D3cold to D0. > > > > > > Say, a PCIe device depending on some ACPI power resources is put into > > > D3cold, so the power resources in question are all turned off. Then, > > > pci_power_up() is called to put it into D0. > > > > > > It first calls pci_platform_power_transition() which invokes > > > platform_pci_set_power_state() to turn on the ACPI power resources > > > depended on by the device and, if that is successful, it calls > > > pci_update_current_state() to update the current_state field of > > > the PCI device object. If the device's configuration space is > > > accessible at this point, which is the case if > > > platform_pci_set_power_state() leaves it in D0-uninitialized (and > > > there's nothing to prevent it from doing so), current_state will be > > > set to PCI_D0 and the pci_raw_set_power_state() called subsequently > > > will notice that the device is in D0 already and do nothing. > > > However, that is not correct, because it may be still necessary to > > > restore the device's BARs at this point. > > > > > > To address this issue, set current_state temporarily to PCI_D3hot > > > in the cases in which pci_raw_set_power_state() may need to do more > > > than just changing the power state of the device. > > > > > > Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > > > > Reviewed-by: Mika Westerberg <mika.westerberg@linux.intel.com> > > Thanks, but on second thought, I'm not sure if this is the best way to > address the issue. > > Basically, pci_power_up() is called in two places, in > pci_set_power_state() (for the transitions to D0) and in > pci_pm_default_resume_early(). In the latter case, > pci_restore_state() is called right after it and that covers BARs > restoration, so nothing more needs to be done in that case. I see. > This means that pci_set_power_state() is the only place needing to > restore the BARs when going into D0 from D3hot or deeper and it is > better to move BARs restoration directly into it. I'll update the > series accordingly and resend. Okay sounds good. > I also think that the mandatory delay is not needed at all when > pci_raw_set_power_state() is called for transitions D3cold -> D0, > because in that case either the device has been powered up via > platform_pci_set_power_state(), or via the bridge resume which takes > the delay into account. I agree.
Index: linux-pm/drivers/pci/pci.c =================================================================== --- linux-pm.orig/drivers/pci/pci.c +++ linux-pm/drivers/pci/pci.c @@ -1309,6 +1309,8 @@ static int pci_dev_wait(struct pci_dev * */ int pci_power_up(struct pci_dev *dev) { + pci_power_t old_state = dev->current_state; + pci_platform_power_transition(dev, PCI_D0); /* @@ -1325,6 +1327,14 @@ int pci_power_up(struct pci_dev *dev) pci_resume_bus(dev->subordinate); } + /* + * For transitions from D3hot or deeper and initial power-up, force + * PCI_PM_CTRL register write, D3*->D0 transition delay and BARS + * restoration. + */ + if (old_state >= PCI_D3hot) + dev->current_state = PCI_D3hot; + return pci_raw_set_power_state(dev, PCI_D0); }