Re: Intel GemniLake xHCI connected devices can never wake up the system from suspend

Re: Intel GemniLake xHCI connected devices can never wake up the system from suspend

[Daniel Drake]

Hi,

I'm working alongside Chris on this issue.

On Fri, Mar 16, 2018 at 12:11 AM, Mathias Nyman
<mathias.nyman@linux.intel.com> wrote:
> Scope (_SB.PCI0.XHC) has _PS0 method, so Linux will look for a _S3W to get
> the
> lowest possible D state in S3, but_S3W is missing, so Linux pci-acpi code
> will
> probably default to D0

Yes, _S3W is missing, and in this case, acpi_dev_pm_get_state() is
setting the maximum number permitted device power state to the minimum
value of D0.

So the XHCI controller is in D0 when you go into suspend.

I tried your hack to force it into D3hot. Now the system can wake up
from resume via the USB port. Thanks for finding that.

>> ASUS said the BIOS has no problem on USB wakeup under Windows so I don't
>> think
>> there's any update. Anything else could be cause for this?
>
> Linux and Windows probably check different DSDT values

I've studied the ACPI spec trying to understand better, but I'm
struggling with the question:
What is the maximum number (lowest power) permitted device power state
for a device that is configured as able to wake the system from S3,
**that does not implement the _S3W method**?

As far as I can see, the ACPI spec doesn't give an answer. It's clear
what the behaviour is when _S3W is present, but not unclear what
should happen when it is not there.

As noted above, Linux's interpretation is that in such case, the
device must remain fully on (D0) when going into S3. I am wondering if
Windows just has made an alternative assumption that all available
device power states can wake the system from suspend in such case.

This is working in Windows so hopefully we can find a way to match the
behaviour?

Thanks
Daniel

Re: Intel GemniLake xHCI connected devices can never wake up the system from suspend

[Daniel Drake]

> I've studied the ACPI spec trying to understand better, but I'm
> struggling with the question:
> What is the maximum number (lowest power) permitted device power state
> for a device that is configured as able to wake the system from S3,
> **that does not implement the _S3W method**?

Actually the ACPI spec has an answer for the case when _S3D is present.
The lack of clarity is only over the situation when both _S3D and _S3W
are missing - like on the platforms being worked on here.

The _S3D docs say:
> If the device can wake the system from the S3 system sleeping state (see
> _PRW) then the device must support wake in the D-state returned by this
> object. However, OSPM cannot assume wake from the S3 system sleeping state
> is supported in any deeper D-state unless specified by a corresponding
> _S3W object

Looking at the design of the existing Linux code, it seems like this
"max = min" assignment that is causing us trouble originates directly
from an attempt to implement that logic: if we didn't get a response from
_S3W, then we must clamp ourselves to the data we got from _S3D.

If I modify the Linux code to be a little more specific in that logic
(only applying when we actually got something from _S3D) then the
problematic behaviour is avoided and USB wakeups work.

I feel that this change makes the Linux implementation more directly
mirror the wording in the ACPI spec and it's associated lack of clarity
for when both methods are missing. Thoughts?

---
 drivers/acpi/device_pm.c | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)

diff --git a/drivers/acpi/device_pm.c b/drivers/acpi/device_pm.c
index a4c8ad98560d..44f12c5c75ee 100644
--- a/drivers/acpi/device_pm.c
+++ b/drivers/acpi/device_pm.c
@@ -543,6 +543,7 @@ static int acpi_dev_pm_get_state(struct device *dev, struct acpi_device *adev,
 	unsigned long long ret;
 	int d_min, d_max;
 	bool wakeup = false;
+	acpi_status sxd_status;
 	acpi_status status;
 
 	/*
@@ -565,8 +566,8 @@ static int acpi_dev_pm_get_state(struct device *dev, struct acpi_device *adev,
 		 * provided if AE_NOT_FOUND is returned.
 		 */
 		ret = d_min;
-		status = acpi_evaluate_integer(handle, method, NULL, &ret);
-		if ((ACPI_FAILURE(status) && status != AE_NOT_FOUND)
+		sxd_status = acpi_evaluate_integer(handle, method, NULL, &ret);
+		if ((ACPI_FAILURE(sxd_status) && sxd_status != AE_NOT_FOUND)
 		    || ret > ACPI_STATE_D3_COLD)
 			return -ENODATA;
 
@@ -599,7 +600,11 @@ static int acpi_dev_pm_get_state(struct device *dev, struct acpi_device *adev,
 		method[3] = 'W';
 		status = acpi_evaluate_integer(handle, method, NULL, &ret);
 		if (status == AE_NOT_FOUND) {
-			if (target_state > ACPI_STATE_S0)
+			/* No _SxW. In this case, the ACPI spec says that we
+			 * must not go into any power state deeper than the
+			 * value returned from _SxD.
+			 */
+			if (sxd_status == AE_OK && target_state > ACPI_STATE_S0)
 				d_max = d_min;
 		} else if (ACPI_SUCCESS(status) && ret <= ACPI_STATE_D3_COLD) {
 			/* Fall back to D3cold if ret is not a valid state. */
-- 
2.14.1

Re: Intel GemniLake xHCI connected devices can never wake up the system from suspend

[Mathias Nyman]

Adding Rafael directly to CC

In short, if _S3D and _S3W are missing in DSDT then a PCI device
stays in D0 during suspend in Linux, but goes to D3 in Windows.

USB wake doesn't work in Geminilake because of this.

Should this be changed? reasoning below.

On 16.03.2018 10:23, Daniel Drake wrote:
>> I've studied the ACPI spec trying to understand better, but I'm
>> struggling with the question:
>> What is the maximum number (lowest power) permitted device power state
>> for a device that is configured as able to wake the system from S3,
>> **that does not implement the _S3W method**?
> 
> Actually the ACPI spec has an answer for the case when _S3D is present.
> The lack of clarity is only over the situation when both _S3D and _S3W
> are missing - like on the platforms being worked on here.
> 
> The _S3D docs say:
>> If the device can wake the system from the S3 system sleeping state (see
>> _PRW) then the device must support wake in the D-state returned by this
>> object. However, OSPM cannot assume wake from the S3 system sleeping state
>> is supported in any deeper D-state unless specified by a corresponding
>> _S3W object
> 
> Looking at the design of the existing Linux code, it seems like this
> "max = min" assignment that is causing us trouble originates directly
> from an attempt to implement that logic: if we didn't get a response from
> _S3W, then we must clamp ourselves to the data we got from _S3D.
> 
> If I modify the Linux code to be a little more specific in that logic
> (only applying when we actually got something from _S3D) then the
> problematic behaviour is avoided and USB wakeups work.
> 
> I feel that this change makes the Linux implementation more directly
> mirror the wording in the ACPI spec and it's associated lack of clarity
> for when both methods are missing. Thoughts?
> 
> ---
>   drivers/acpi/device_pm.c | 11 ++++++++---
>   1 file changed, 8 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/acpi/device_pm.c b/drivers/acpi/device_pm.c
> index a4c8ad98560d..44f12c5c75ee 100644
> --- a/drivers/acpi/device_pm.c
> +++ b/drivers/acpi/device_pm.c
> @@ -543,6 +543,7 @@ static int acpi_dev_pm_get_state(struct device *dev, struct acpi_device *adev,
>   	unsigned long long ret;
>   	int d_min, d_max;
>   	bool wakeup = false;
> +	acpi_status sxd_status;
>   	acpi_status status;
>   
>   	/*
> @@ -565,8 +566,8 @@ static int acpi_dev_pm_get_state(struct device *dev, struct acpi_device *adev,
>   		 * provided if AE_NOT_FOUND is returned.
>   		 */
>   		ret = d_min;
> -		status = acpi_evaluate_integer(handle, method, NULL, &ret);
> -		if ((ACPI_FAILURE(status) && status != AE_NOT_FOUND)
> +		sxd_status = acpi_evaluate_integer(handle, method, NULL, &ret);
> +		if ((ACPI_FAILURE(sxd_status) && sxd_status != AE_NOT_FOUND)
>   		    || ret > ACPI_STATE_D3_COLD)
>   			return -ENODATA;
>   
> @@ -599,7 +600,11 @@ static int acpi_dev_pm_get_state(struct device *dev, struct acpi_device *adev,
>   		method[3] = 'W';
>   		status = acpi_evaluate_integer(handle, method, NULL, &ret);
>   		if (status == AE_NOT_FOUND) {
> -			if (target_state > ACPI_STATE_S0)
> +			/* No _SxW. In this case, the ACPI spec says that we
> +			 * must not go into any power state deeper than the
> +			 * value returned from _SxD.
> +			 */
> +			if (sxd_status == AE_OK && target_state > ACPI_STATE_S0)
>   				d_max = d_min;
>   		} else if (ACPI_SUCCESS(status) && ret <= ACPI_STATE_D3_COLD) {
>   			/* Fall back to D3cold if ret is not a valid state. */
> 

Re: Intel GemniLake xHCI connected devices can never wake up the system from suspend

[Rafael J. Wysocki]

On Fri, Mar 16, 2018 at 10:06 AM, Mathias Nyman
<mathias.nyman@linux.intel.com> wrote:
> Adding Rafael directly to CC
>
> In short, if _S3D and _S3W are missing in DSDT then a PCI device
> stays in D0 during suspend in Linux, but goes to D3 in Windows.
>
> USB wake doesn't work in Geminilake because of this.
>
> Should this be changed? reasoning below.

It can be changed if that doesn't cause problems to happen.


> On 16.03.2018 10:23, Daniel Drake wrote:
>>>
>>> I've studied the ACPI spec trying to understand better, but I'm
>>> struggling with the question:
>>> What is the maximum number (lowest power) permitted device power state
>>> for a device that is configured as able to wake the system from S3,
>>> **that does not implement the _S3W method**?
>>
>>
>> Actually the ACPI spec has an answer for the case when _S3D is present.
>> The lack of clarity is only over the situation when both _S3D and _S3W
>> are missing - like on the platforms being worked on here.
>>
>> The _S3D docs say:
>>>
>>> If the device can wake the system from the S3 system sleeping state (see
>>> _PRW) then the device must support wake in the D-state returned by this
>>> object. However, OSPM cannot assume wake from the S3 system sleeping
>>> state
>>> is supported in any deeper D-state unless specified by a corresponding
>>> _S3W object
>>
>>
>> Looking at the design of the existing Linux code, it seems like this
>> "max = min" assignment that is causing us trouble originates directly
>> from an attempt to implement that logic: if we didn't get a response from
>> _S3W, then we must clamp ourselves to the data we got from _S3D.
>>
>> If I modify the Linux code to be a little more specific in that logic
>> (only applying when we actually got something from _S3D) then the
>> problematic behaviour is avoided and USB wakeups work.
>>
>> I feel that this change makes the Linux implementation more directly
>> mirror the wording in the ACPI spec and it's associated lack of clarity
>> for when both methods are missing. Thoughts?
>>
>> ---
>>   drivers/acpi/device_pm.c | 11 ++++++++---
>>   1 file changed, 8 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/acpi/device_pm.c b/drivers/acpi/device_pm.c
>> index a4c8ad98560d..44f12c5c75ee 100644
>> --- a/drivers/acpi/device_pm.c
>> +++ b/drivers/acpi/device_pm.c
>> @@ -543,6 +543,7 @@ static int acpi_dev_pm_get_state(struct device *dev,
>> struct acpi_device *adev,
>>         unsigned long long ret;
>>         int d_min, d_max;
>>         bool wakeup = false;
>> +       acpi_status sxd_status;
>>         acpi_status status;
>>         /*
>> @@ -565,8 +566,8 @@ static int acpi_dev_pm_get_state(struct device *dev,
>> struct acpi_device *adev,
>>                  * provided if AE_NOT_FOUND is returned.
>>                  */
>>                 ret = d_min;
>> -               status = acpi_evaluate_integer(handle, method, NULL,
>> &ret);
>> -               if ((ACPI_FAILURE(status) && status != AE_NOT_FOUND)
>> +               sxd_status = acpi_evaluate_integer(handle, method, NULL,
>> &ret);
>> +               if ((ACPI_FAILURE(sxd_status) && sxd_status !=
>> AE_NOT_FOUND)
>>                     || ret > ACPI_STATE_D3_COLD)
>>                         return -ENODATA;
>>   @@ -599,7 +600,11 @@ static int acpi_dev_pm_get_state(struct device
>> *dev, struct acpi_device *adev,
>>                 method[3] = 'W';
>>                 status = acpi_evaluate_integer(handle, method, NULL,
>> &ret);
>>                 if (status == AE_NOT_FOUND) {
>> -                       if (target_state > ACPI_STATE_S0)
>> +                       /* No _SxW. In this case, the ACPI spec says that
>> we
>> +                        * must not go into any power state deeper than
>> the
>> +                        * value returned from _SxD.
>> +                        */
>> +                       if (sxd_status == AE_OK && target_state >
>> ACPI_STATE_S0)
>>                                 d_max = d_min;
>>                 } else if (ACPI_SUCCESS(status) && ret <=
>> ACPI_STATE_D3_COLD) {
>>                         /* Fall back to D3cold if ret is not a valid
>> state. */
>>
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-acpi" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: Intel GemniLake xHCI connected devices can never wake up the system from suspend

[Daniel Drake]

On Mon, Mar 19, 2018 at 6:36 AM, Rafael J. Wysocki <rafael@kernel.org> wrote:
> On Fri, Mar 16, 2018 at 10:06 AM, Mathias Nyman
> <mathias.nyman@linux.intel.com> wrote:
>> Adding Rafael directly to CC
>>
>> In short, if _S3D and _S3W are missing in DSDT then a PCI device
>> stays in D0 during suspend in Linux, but goes to D3 in Windows.
>>
>> USB wake doesn't work in Geminilake because of this.
>>
>> Should this be changed? reasoning below.
>
> It can be changed if that doesn't cause problems to happen.

I double checked that Windows 10 is going into S3 suspend and that USB
wakeup works fine on this platform - it works fine there.
Device manager for the XHCI controller clearly shows D3 being used for
S3 suspend: https://imgur.com/lF9U3V0

Current power state:
D0

Power capabilities:
00000089
PDCAP_D0_SUPPORTED
PDCAP_D3_SUPPORTED
PDCAP_WAKE_FROM_D3_SUPPORTED

Power state mappings:
S0 -> D0
S1 -> Unspecified
S2 -> Unspecified
S3 -> D3
S4 -> D3
S5 -> D3

The biggest risk of my proposed change is that we would now end up
putting a wakeup-enabled device into a too low power state where it
can no longer wake up the system. On the other hand, it solves this
issue affecting 2 different vendors, and may even result in some power
savings in general.

Perhaps we can consider the change for inclusion in Linux-4.17, giving
2-3 months of testing time.

I'll submit the patch shortly.

Thanks
Daniel