USB hot-plug not working (ASUS TP301UA-C4028T)

Quoted 52 lines · Alan Stern

Quoted 37 lines · Mathias Nyman

Quoted 5 lines · Pierre de Villemereuil

Hi Mathias,

Just to know: does that mean the firmware from Asus is faulty in here? Do you think I should contact Asus about this?

Probably, _S0W, _DSM and XFLT code for xHCI are useless in that ACPI DSDT firmware version.

But we still want the driver to handle cases like this. Just wrote the patch.
Adding Alan Stern to CC for PM sanity feedback on it:

---

Author: Mathias Nyman <[email protected]>
Date:   Tue Oct 4 13:07:59 2016 +0300

xhci: Prevent a non-responsive xhci suspend state.

ACPI may limit the lowest possible runtime suspend PCI D-state to D0. If xHC is not capable of generating PME# events in D0 we end up with a non-responsive runtime suspended host without PME# capability and with interrupts disabled, unable to detect or wake up when a new usb device is connected.

This was triggered on a ASUS TP301UA machine.

Reported-by Pierre de Villemereuil <[email protected]>

Signed-off-by Mathias Nyman <[email protected]>

drivers/usb/host/xhci-pci.c +0 -0

index d7b0f97..08b021c 100644
@@ -396,6 +396,11 @@ static int xhci_pci_suspend(struct usb_hcd *hcd, bool do_wakeup)
          if (xhci->quirks & XHCI_SSIC_PORT_UNUSED)

drivers/usb/host/xhci-pci.c static int xhci_pci_suspend(struct usb_hcd *hcd, bool do_wakeup) · line 406

Quoted 10 lines

drivers/usb/host/xhci-pci.c +5 -0

                  xhci_ssic_port_unused_quirk(hcd, true);
 
+       /* Prevent a non-responsive PCI D0 state without PME# wake capability */
+       if (pci_choose_state(pdev, PMSG_AUTO_SUSPEND) == PCI_D0)
+               if (!pci_pme_capable(pdev, PCI_D0))
+                       return -EBUSY;
+
          ret = xhci_suspend(xhci, do_wakeup);
          if (ret && (xhci->quirks & XHCI_SSIC_PORT_UNUSED))
                  xhci_ssic_port_unused_quirk(hcd, false);

I've got three comments about this proposal.

Quoted 3 lines

First, this logic should not apply during system suspend, only during runtime suspend. You don't want to abort putting a laptop to sleep just because the xHCI controller can't generate wakeup signals.

Yes, I assume it would be ok change usb core to pass down something like pm_message_t to suspend_common() so that we could do this. hcd_pci_runetime_suspend() -> suspend_common() -> hcd->driver->pci_suspend()

Assuming this workaround should stay in xhci-pci.c

Quoted 4 lines

Second, this really has nothing to do with the D0 state. The same logic should apply to whatever state is chosen for runtime suspend: If the controller can't generate PME# wakeup signals in that state then the suspend should be aborted.

PCI code actually does this for us, it will walk down the D-states until it finds one that support PME#, but stop at D0 end return D0 even if D0 does not support PME#.

Unfortunately that is done in a static function in pci/pci.c.

static pci_power_t pci_target_state(struct pci_dev *dev)

{
         pci_power_t target_state = PCI_D3hot;

	if (platform_pci_power_manageable(dev)) {
                 /*
                  * Call the platform to choose the target state of the device
                  * and enable wake-up from this state if supported.
                  */
                 pci_power_t state = platform_pci_choose_state(dev);

	        switch (state) {
	        case PCI_POWER_ERROR:
                 case PCI_UNKNOWN:
                         break;
                 case PCI_D1:
                 case PCI_D2:
                         if (pci_no_d1d2(dev))
	                        break;
                 default:
	                target_state = state;
                 }
         } else if (!dev->pm_cap) {
                 target_state = PCI_D0;
         } else if (device_may_wakeup(&dev->dev)) {
                 /*
                  * Find the deepest state from which the device can generate
                  * wake-up events, make it the target state and enable device
                  * to generate PME#.
                  */
                 if (dev->pme_support) {
                         while (target_state
                               && !(dev->pme_support & (1 << target_state)))
                                 target_state--;
                 }
         }

         return target_state;
}

The best I can do from xhci is call pci_choose_state() which will call platform_pci_choose_state(), but won't do the PME# checking and D-state decrement to D0 .

If pci_choose_state() returns D0 from a runtime suspend callback (and yes, should make sure its runtime suspend, not system suspend) I can pretty much assume pci_target_state will do the same.

Quoted 5 lines

Third, the same reasoning applies to every PCI device that relies on PME#, not just to xHCI controllers. Therefore the new code should be added to drivers/pci/pci-driver.c:pci_pm_runtime_suspend(), not to xhci-pci.c.

Yes, that would be the preferred solution.
I was not sure we can do that. pci-driver.c pci_pm_runtime_suspend() will call pci_finish_runtime_suspend()
pci_target_state() // returns D0,
pci_set_power_state() // to D0, which is the state we probably are already in.

So it won't really do much. The device PCI device is disabled in usb hcd suspend hcd-pci.c:
suspend_common()
...
pci_disable_device(pci_dev)

Should we anyway disable runtime PM in PCI if the PCI device has wakeup enabled and the target state is D0 without PME# support?

[Adding Bjorn and linux-pci to the CC: list]

In short, Pierre's USB host controller doesn't send wakeup signals from runtime suspend, because the firmware limits the runtime-suspend state to D0 and the controller can't issue PME# from the D0 state. In this situation we would prefer to avoid suspending the controller at all, rather than have it go into runtime suspend and then stop working.

Quoted 57 lines · Mathias Nyman

Quoted 52 lines · Alan Stern

Quoted 37 lines · Mathias Nyman

Quoted 5 lines · Pierre de Villemereuil

Hi Mathias,

Just to know: does that mean the firmware from Asus is faulty in here? Do you think I should contact Asus about this?

Probably, _S0W, _DSM and XFLT code for xHCI are useless in that ACPI DSDT firmware version.

But we still want the driver to handle cases like this. Just wrote the patch.
Adding Alan Stern to CC for PM sanity feedback on it:

---

Author: Mathias Nyman <[email protected]>
Date:   Tue Oct 4 13:07:59 2016 +0300

xhci: Prevent a non-responsive xhci suspend state.

ACPI may limit the lowest possible runtime suspend PCI D-state to D0. If xHC is not capable of generating PME# events in D0 we end up with a non-responsive runtime suspended host without PME# capability and with interrupts disabled, unable to detect or wake up when a new usb device is connected.

This was triggered on a ASUS TP301UA machine.

Reported-by Pierre de Villemereuil <[email protected]>

Signed-off-by Mathias Nyman <[email protected]>

drivers/usb/host/xhci-pci.c +0 -0

index d7b0f97..08b021c 100644
@@ -396,6 +396,11 @@ static int xhci_pci_suspend(struct usb_hcd *hcd, bool do_wakeup)
          if (xhci->quirks & XHCI_SSIC_PORT_UNUSED)

drivers/usb/host/xhci-pci.c static int xhci_pci_suspend(struct usb_hcd *hcd, bool do_wakeup) · line 406

Quoted 10 lines

drivers/usb/host/xhci-pci.c +5 -0

                  xhci_ssic_port_unused_quirk(hcd, true);
 
+       /* Prevent a non-responsive PCI D0 state without PME# wake capability */
+       if (pci_choose_state(pdev, PMSG_AUTO_SUSPEND) == PCI_D0)
+               if (!pci_pme_capable(pdev, PCI_D0))
+                       return -EBUSY;
+
          ret = xhci_suspend(xhci, do_wakeup);
          if (ret && (xhci->quirks & XHCI_SSIC_PORT_UNUSED))
                  xhci_ssic_port_unused_quirk(hcd, false);

I've got three comments about this proposal.

Quoted 3 lines

First, this logic should not apply during system suspend, only during runtime suspend. You don't want to abort putting a laptop to sleep just because the xHCI controller can't generate wakeup signals.

Quoted 5 lines

Yes, I assume it would be ok change usb core to pass down something like pm_message_t to suspend_common() so that we could do this. hcd_pci_runetime_suspend() -> suspend_common() -> hcd->driver->pci_suspend()

Assuming this workaround should stay in xhci-pci.c

If necessary, it could be moved into hcd_pci_runtime_suspend(). But I would prefer to put it in the PCI core.

Quoted 8 lines

Quoted 4 lines

Second, this really has nothing to do with the D0 state. The same logic should apply to whatever state is chosen for runtime suspend: If the controller can't generate PME# wakeup signals in that state then the suspend should be aborted.

PCI code actually does this for us, it will walk down the D-states until it finds one that support PME#, but stop at D0 end return D0 even if D0 does not support PME#.

That sounds like a bug. Or rather, accepting D0 as the target state when it doesn't support PME# is the bug.

Quoted 62 lines

Unfortunately that is done in a static function in pci/pci.c.

static pci_power_t pci_target_state(struct pci_dev *dev)

{
         pci_power_t target_state = PCI_D3hot;

	if (platform_pci_power_manageable(dev)) {
                 /*
                  * Call the platform to choose the target state of the device
                  * and enable wake-up from this state if supported.
                  */
                 pci_power_t state = platform_pci_choose_state(dev);

	        switch (state) {
	        case PCI_POWER_ERROR:
                 case PCI_UNKNOWN:
                         break;
                 case PCI_D1:
                 case PCI_D2:
                         if (pci_no_d1d2(dev))
	                        break;
                 default:
	                target_state = state;
                 }
         } else if (!dev->pm_cap) {
                 target_state = PCI_D0;
         } else if (device_may_wakeup(&dev->dev)) {
                 /*
                  * Find the deepest state from which the device can generate
                  * wake-up events, make it the target state and enable device
                  * to generate PME#.
                  */
                 if (dev->pme_support) {
                         while (target_state
                               && !(dev->pme_support & (1 << target_state)))
                                 target_state--;
                 }
         }

         return target_state;
}

The best I can do from xhci is call pci_choose_state() which will call platform_pci_choose_state(), but won't do the PME# checking and D-state decrement to D0 .

If pci_choose_state() returns D0 from a runtime suspend callback (and yes, should make sure its runtime suspend, not system suspend) I can pretty much assume pci_target_state will do the same.

Quoted 5 lines

Third, the same reasoning applies to every PCI device that relies on PME#, not just to xHCI controllers. Therefore the new code should be added to drivers/pci/pci-driver.c:pci_pm_runtime_suspend(), not to xhci-pci.c.

Yes, that would be the preferred solution.
I was not sure we can do that. pci-driver.c pci_pm_runtime_suspend() will call pci_finish_runtime_suspend()
pci_target_state() // returns D0,
pci_set_power_state() // to D0, which is the state we probably are already in.

Quoted 8 lines

So it won't really do much. The device PCI device is disabled in usb hcd suspend hcd-pci.c:
suspend_common()
...
pci_disable_device(pci_dev)

Should we anyway disable runtime PM in PCI if the PCI device has wakeup enabled and the target state is D0 without PME# support?

First, the device_may_wakeup() test applies only to suspend suspend, not to runtime suspend. The current PM design assumes that wakeup will always be enabled during runtime suspend (if the device supports it).

Second, there is a potential problem because some PCI devices have other platform-based mechanisms for sending wakeup signals, instead of using PME#. Such devices probably don't support standard PCI wakeup at all. (A good example is the UHCI controllers on Intel's old chipsets.) We don't want to prevent them from going into runtime suspend.

Now, maybe this isn't an issue -- I don't know enough about the details of how the PCI core handles runtime suspend and how it coordinates with the platform code.

Hopefully Bjorn can fill in the missing details.

Alan Stern

--
To unsubscribe from this list: send the line "unsubscribe linux-pci" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

Quoted 5 lines · Alan Stern

In short, Pierre's USB host controller doesn't send wakeup signals from runtime suspend, because the firmware limits the runtime-suspend state to D0 and the controller can't issue PME# from the D0 state. In this situation we would prefer to avoid suspending the controller at all, rather than have it go into runtime suspend and then stop working.

Quoted 169 lines · Mathias Nyman

Quoted 58 lines · Alan Stern

Quoted 52 lines · Mathias Nyman

Quoted 38 lines · Pierre de Villemereuil

Quoted 5 lines

Hi Mathias,

Just to know: does that mean the firmware from Asus is faulty in here? Do you think I should contact Asus about this?

Probably, _S0W, _DSM and XFLT code for xHCI are useless in that ACPI DSDT firmware version.

But we still want the driver to handle cases like this. Just wrote the patch.
Adding Alan Stern to CC for PM sanity feedback on it:

---

Author: Mathias Nyman <[email protected]>
Date:   Tue Oct 4 13:07:59 2016 +0300

xhci: Prevent a non-responsive xhci suspend state.

ACPI may limit the lowest possible runtime suspend PCI D-state to D0. If xHC is not capable of generating PME# events in D0 we end up with a non-responsive runtime suspended host without PME# capability and with interrupts disabled, unable to detect or wake up when a new usb device is connected.

This was triggered on a ASUS TP301UA machine.

Reported-by Pierre de Villemereuil <[email protected]>

Signed-off-by Mathias Nyman <[email protected]>

drivers/usb/host/xhci-pci.c +0 -0

index d7b0f97..08b021c 100644
@@ -396,6 +396,11 @@ static int xhci_pci_suspend(struct usb_hcd *hcd, bool do_wakeup)
          if (xhci->quirks & XHCI_SSIC_PORT_UNUSED)

drivers/usb/host/xhci-pci.c static int xhci_pci_suspend(struct usb_hcd *hcd, bool do_wakeup) · line 406

Quoted 10 lines

drivers/usb/host/xhci-pci.c +5 -0

                  xhci_ssic_port_unused_quirk(hcd, true);
 
+       /* Prevent a non-responsive PCI D0 state without PME# wake capability */
+       if (pci_choose_state(pdev, PMSG_AUTO_SUSPEND) == PCI_D0)
+               if (!pci_pme_capable(pdev, PCI_D0))
+                       return -EBUSY;
+
          ret = xhci_suspend(xhci, do_wakeup);
          if (ret && (xhci->quirks & XHCI_SSIC_PORT_UNUSED))
                  xhci_ssic_port_unused_quirk(hcd, false);

I've got three comments about this proposal.

Quoted 3 lines

First, this logic should not apply during system suspend, only during runtime suspend. You don't want to abort putting a laptop to sleep just because the xHCI controller can't generate wakeup signals.

Quoted 5 lines

Yes, I assume it would be ok change usb core to pass down something like pm_message_t to suspend_common() so that we could do this. hcd_pci_runetime_suspend() -> suspend_common() -> hcd->driver->pci_suspend()

Assuming this workaround should stay in xhci-pci.c

If necessary, it could be moved into hcd_pci_runtime_suspend(). But I would prefer to put it in the PCI core.

Quoted 8 lines

Quoted 4 lines

Second, this really has nothing to do with the D0 state. The same logic should apply to whatever state is chosen for runtime suspend: If the controller can't generate PME# wakeup signals in that state then the suspend should be aborted.

PCI code actually does this for us, it will walk down the D-states until it finds one that support PME#, but stop at D0 end return D0 even if D0 does not support PME#.

That sounds like a bug. Or rather, accepting D0 as the target state when it doesn't support PME# is the bug.

Quoted 62 lines

Unfortunately that is done in a static function in pci/pci.c.

static pci_power_t pci_target_state(struct pci_dev *dev)

{
         pci_power_t target_state = PCI_D3hot;

	if (platform_pci_power_manageable(dev)) {
                 /*
                  * Call the platform to choose the target state of the device
                  * and enable wake-up from this state if supported.
                  */
                 pci_power_t state = platform_pci_choose_state(dev);

	        switch (state) {
	        case PCI_POWER_ERROR:
                 case PCI_UNKNOWN:
                         break;
                 case PCI_D1:
                 case PCI_D2:
                         if (pci_no_d1d2(dev))
	                        break;
                 default:
	                target_state = state;
                 }
         } else if (!dev->pm_cap) {
                 target_state = PCI_D0;
         } else if (device_may_wakeup(&dev->dev)) {
                 /*
                  * Find the deepest state from which the device can generate
                  * wake-up events, make it the target state and enable device
                  * to generate PME#.
                  */
                 if (dev->pme_support) {
                         while (target_state
                               && !(dev->pme_support & (1 << target_state)))
                                 target_state--;
                 }
         }

         return target_state;
}

The best I can do from xhci is call pci_choose_state() which will call platform_pci_choose_state(), but won't do the PME# checking and D-state decrement to D0 .

If pci_choose_state() returns D0 from a runtime suspend callback (and yes, should make sure its runtime suspend, not system suspend) I can pretty much assume pci_target_state will do the same.

Quoted 5 lines

Third, the same reasoning applies to every PCI device that relies on PME#, not just to xHCI controllers. Therefore the new code should be added to drivers/pci/pci-driver.c:pci_pm_runtime_suspend(), not to xhci-pci.c.

Yes, that would be the preferred solution.
I was not sure we can do that. pci-driver.c pci_pm_runtime_suspend() will call pci_finish_runtime_suspend()
pci_target_state() // returns D0,
pci_set_power_state() // to D0, which is the state we probably are already in.

Quoted 8 lines

So it won't really do much. The device PCI device is disabled in usb hcd suspend hcd-pci.c:
suspend_common()
...
pci_disable_device(pci_dev)

Should we anyway disable runtime PM in PCI if the PCI device has wakeup enabled and the target state is D0 without PME# support?

First, the device_may_wakeup() test applies only to suspend suspend, not to runtime suspend. The current PM design assumes that wakeup will always be enabled during runtime suspend (if the device supports it).

Second, there is a potential problem because some PCI devices have other platform-based mechanisms for sending wakeup signals, instead of using PME#. Such devices probably don't support standard PCI wakeup at all. (A good example is the UHCI controllers on Intel's old chipsets.) We don't want to prevent them from going into runtime suspend.

Now, maybe this isn't an issue -- I don't know enough about the details of how the PCI core handles runtime suspend and how it coordinates with the platform code.

Hopefully Bjorn can fill in the missing details.

Alan Stern

The PCI core doesn't allow runtime PM by default. Rather it calls pm_runtime_forbid() when the device is added (see pci_pm_init(), called
indirectly from pci_device_add()). PCI drivers need to explicitly call
pm_runtime_allow(), typically from their ->probe hook.

If this xHC cannot signal wakeup, it shouldn't allow runtime PM in the first place. Simple as that.

Quoted 6 lines · Bjorn Helgaas

Quoted 5 lines · Alan Stern

In short, Pierre's USB host controller doesn't send wakeup signals from runtime suspend, because the firmware limits the runtime-suspend state to D0 and the controller can't issue PME# from the D0 state. In this situation we would prefer to avoid suspending the controller at all, rather than have it go into runtime suspend and then stop working.

As Alan has correctly pointed out below, there are PCI devices which do not support PME but should still be runtime suspended, e.g. because they have some non-standard mechanism to sideband signal wakeup or because they can detect when they need to resume even if they're in a low-power state.

AFAIUI, this device should not be runtime suspended at all because it doesn't generate a PME interrupt and thus stays suspended forever.

The PCI core doesn't allow runtime PM by default. Rather it calls pm_runtime_forbid() when the device is added (see pci_pm_init(), called indirectly from pci_device_add()). PCI drivers need to explicitly call pm_runtime_allow(), typically from their ->probe hook.

If this xHC cannot signal wakeup, it shouldn't allow runtime PM in the first place. Simple as that.

The USB core has a function usb_enable_autosuspend() which does nothing else but call pm_runtime_allow(). This is called in a couple of places, I guess the relevant caller here is drivers/usb/core/hub.c:hub_probe(). I'd suggest to amend that function so that runtime PM isn't allowed if the host controller can't signal plug events from pci_target_state().

This approach is better than returning -EBUSY from the ->runtime_suspend hook because the PM core doesn't waste CPU cycles by repeatedly calling ->runtime_suspend in vain, always getting -EBUSY back.

Quoted 14 lines · Lukas Wunner

Quoted 6 lines · Bjorn Helgaas

Quoted 5 lines · Alan Stern

In short, Pierre's USB host controller doesn't send wakeup signals from runtime suspend, because the firmware limits the runtime-suspend state to D0 and the controller can't issue PME# from the D0 state. In this situation we would prefer to avoid suspending the controller at all, rather than have it go into runtime suspend and then stop working.

As Alan has correctly pointed out below, there are PCI devices which do not support PME but should still be runtime suspended, e.g. because they have some non-standard mechanism to sideband signal wakeup or because they can detect when they need to resume even if they're in a low-power state.

Quoted 2 lines

AFAIUI, this device should not be runtime suspended at all because it doesn't generate a PME interrupt and thus stays suspended forever.

No, as Oliver said, the device can generate a PME# signal. It just can't do so from D0, and the firmware doesn't allow it to go into a deeper power-savings state.

Quoted 5 lines

The PCI core doesn't allow runtime PM by default. Rather it calls pm_runtime_forbid() when the device is added (see pci_pm_init(), called indirectly from pci_device_add()). PCI drivers need to explicitly call pm_runtime_allow(), typically from their ->probe hook.

No, pm_runtime_allow() is generally called by userspace, via writing to the .../power/control file in sysfs. Most drivers do not use it; it is a policy mechanism. And drivers can't use it to _enforce_ anything, since the user can always override the setting.

Quoted 4 lines

If this xHC cannot signal wakeup, it shouldn't allow runtime PM in the first place. Simple as that.

I still think this belongs in the PCI core -- except for the difficulty of determining whether a device can use a non-PME method for wakeup signalling. If that issue has a good solution then the PCI core could call pm_runtime_get_noresume() for devices that are capable of generating wakeup signals but not in any D-state that the firmware will allow for runtime suspend.

The PCI core already calls pm_runtime_get_sync() before invoking the ->probe hook of a driver (see local_pci_probe()). Drivers need to explicitly release a runtime ref to allow their device to suspend. For xhci-pci, this seems to happen in usb_hcd_pci_probe():

	if (pci_dev_run_wake(dev))
		pm_runtime_put_noidle(&dev->dev);

So you could either modify the if-condition if you want to change the behaviour for XHCI devices only, or if you want to change it in general, add something like this to pci_dev_run_wake():

	/* PME capable in principle, but not from the intended sleep state */
	if (dev->pme_support && !pci_pme_capable(dev, pci_target_state(dev)))
		return false;

I've briefly looked over the callers of pci_dev_run_wake() and the above seems safe but you should double-check them.

Quoted 1 line

The PCI core already calls pm_runtime_get_sync() before invoking=20

Quoted 1 line

->probe hook of a driver (see local_pci_probe()). Drivers need=20

Quoted 1 line

explicitly release a runtime ref to allow their device to=20

Quoted 7 lines

=20
For xhci-pci, this seems to happen in usb_hcd_pci_probe():

	if (pci_dev_run_wake(dev))
=09
		pm_runtime_put_noidle(&dev->dev);

=20
So you could either modify the if-condition if you want to=20

Quoted 1 line

behaviour for XHCI devices only, or if you want to change it in=20

Quoted 3 lines

=20
add something like this to pci_dev_run_wake():

	/* PME capable in principle, but not from the intended sleep=20

Quoted 1 line

	if (dev->pme_support && !pci_pme_capable(dev,=20

Quoted 4 lines

=09

		return false;

=20
I've briefly looked over the callers of pci_dev_run_wake() and=20

Quoted 1 line

seems safe but you should double-check them.

=20
That seems like a good suggestion. The patch is below; Pierre,=20

test it? This should remove the need to set the USB autosuspend=20

parameter to -1.
=20
Alan Stern
=20
=20
=20
Index: usb-4.x/drivers/pci/pci.c
=20

drivers/pci/pci.c +3 -0

+	if (!pci_pme_capable(dev, pci_target_state(dev)))
+		return false;
+
 	while (bus->parent) {
 		struct pci_dev *bridge =3D bus->self;

Hi!

I'm sorry, I'm not savvy enough to know what to do with this (I know basics on how to code and compile, but I'm no dev). Could someone guide me through it?

I gather this patch needs to be applied to some kernel module code which needs to be compiled and reloaded into my current kernel, right?

Quoted 3 lines

Hi!
=20
I'm sorry, I'm not savvy enough to know what to do with this (I=20

Quoted 1 line

basics on how to code and compile, but I'm no dev). Could=20

Quoted 3 lines

guide me through it?
=20
I gather this patch needs to be applied to some kernel module=20

Quoted 2 lines

which needs to be compiled and reloaded into my current kernel, right?

=20
More likely you'll have to rebuild an entire new kernel, not just=20

single module.
=20
Search the support site for the distribution you are using. It=20

to contain reasonably thorough instructions on how to build and=20

your own version of their kernel.
=20
Once you know how to do all that, I can tell you how the patch=20

be applied -- that's the easy part.
=20
Alan Stern

Hi!

I've branched the kernel package on the OpenSUSE Build Server, I'll try to apply the patch there (this ought to be cleanest method).

Starting from the root of the kernel tarball, the patch should be applied to drivers/pci/pci.c, am I right?

Quoted 3 lines

Hi!
=20
I've branched the kernel package on the OpenSUSE Build Server,=20

Quoted 3 lines

try to apply the patch there (this ought to be cleanest method). =20
Starting from the root of the kernel tarball, the patch should=20

Quoted 1 line

applied to drivers/pci/pci.c, am I right?

=20
You just cd to the top directory of the kernel source, and do =20
patch -p1 <filename
=20
where "filename" is the contents of the email message containing=20

patch. Or if you want, since the patch is so small, you can=20

edit the file by hand to add in the new lines.
=20
Alan Stern

Quoted 16 lines · Lukas Wunner

The PCI core already calls pm_runtime_get_sync() before invoking the ->probe hook of a driver (see local_pci_probe()). Drivers need to explicitly release a runtime ref to allow their device to suspend. For xhci-pci, this seems to happen in usb_hcd_pci_probe():

	if (pci_dev_run_wake(dev))
		pm_runtime_put_noidle(&dev->dev);

So you could either modify the if-condition if you want to change the behaviour for XHCI devices only, or if you want to change it in general, add something like this to pci_dev_run_wake():

	/* PME capable in principle, but not from the intended sleep state */
	if (dev->pme_support && !pci_pme_capable(dev, pci_target_state(dev)))
		return false;

Quoted 2 lines

I've briefly looked over the callers of pci_dev_run_wake() and the above seems safe but you should double-check them.

That seems like a good suggestion. The patch is below; Pierre, can you test it? This should remove the need to set the USB autosuspend module parameter to -1.

Index: usb-4.x/drivers/pci/pci.c
===================================================================

drivers/pci/pci.c +4 -0

@@ -2064,6 +2064,10 @@ bool pci_dev_run_wake(struct pci_dev *de
 	if (!dev->pme_support)
 		return false;
 
+	/* PME-capable in principle, but not from the intended sleep state */
+	if (!pci_pme_capable(dev, pci_target_state(dev)))
+		return false;
+
 	while (bus->parent) {
 		struct pci_dev *bridge = bus->self;
 

Quoted 20 lines · Alan Stern

Quoted 18 lines · Lukas Wunner

The PCI core already calls pm_runtime_get_sync() before invoking the ->probe hook of a driver (see local_pci_probe()). Drivers need to explicitly release a runtime ref to allow their device to suspend. For xhci-pci, this seems to happen in usb_hcd_pci_probe():

	if (pci_dev_run_wake(dev))
		pm_runtime_put_noidle(&dev->dev);

So you could either modify the if-condition if you want to change the behaviour for XHCI devices only, or if you want to change it in general, add something like this to pci_dev_run_wake():

	/* PME capable in principle, but not from the intended sleep state */
	if (dev->pme_support && !pci_pme_capable(dev, pci_target_state(dev)))
		return false;

I've briefly looked over the callers of pci_dev_run_wake() and the above seems safe but you should double-check them.

Quoted 3 lines

That seems like a good suggestion. The patch is below; Pierre, can you test it? This should remove the need to set the USB autosuspend module parameter to -1.

Alan, how do we proceed with this? Are you going to submit a patch (with commit message, tags and all) to linux-pci@ or would you prefer me to do that? I just went over the callers of pci_dev_run_wake() once more and the patch still looks safe to me.