PCI / ACPI: Handle sibling devices sharing power resources
Subsystems (3)
v2 cover PCI / ACPI: Handle sibling devices sharing power resources
Hi all,
Based on a discussion regarding patch series I sent previously [1] to deal with sibling devices sharing ACPI power resources, I prepared a new reworked version according to the comments I got.
To summarize, in Intel Ice Lake the Thunderbolt controller, PCIe root ports and xHCI all share power resources. When they are all in D3hot power resources (returned by _PR3) can be turned off powering off the whole block. However, there are two issues around this.
Firstly the PCI core sets the device power state by asking what the real ACPI power state is. This results that all but last device sharing the power resources are in D3hot when the power resources are turned off. This causes issues if user runs for example 'lspci' because the device is really in D3cold so what user gets back is all ones (0xffffffff).
Secondly if any of the device is runtime resumed the power resources are turned on bringing all other devices sharing the resources to D0uninitialized losing their wakeup configuration.
This series aims to fix the two issues by:
1. Using the ACPI cached power state when PCI devices are transitioned into low power states instead of reading back the "real" power state.
2. Introducing concept of "_PR0 dependent devices" that get runtime resumed whenever their power resource (which they might share with other sibling devices) gets turned on.
The series is based on the idea of Rafael J. Wysocki <[email protected]>.
[1] https://www.spinics.net/lists/linux-pci/msg83583.html
Mika Westerberg (3):
PCI / ACPI: Use cached ACPI device state to get PCI device power state ACPI / PM: Introduce concept of a _PR0 dependent device PCI / ACPI: Add _PR0 dependent devices
drivers/acpi/power.c | 139 ++++++++++++++++++++++++++++++++++++++++ drivers/pci/pci-acpi.c | 5 +- include/acpi/acpi_bus.h | 4 ++ 3 files changed, 147 insertions(+), 1 deletion(-)
Signature · 1 line
2.20.1
PATCH v2 1/3 PCI / ACPI: Use cached ACPI device state to get PCI device power state
Intel Ice Lake has an integrated Thunderbolt controller which means that the PCIe topology is extended directly from the two root ports (RP0 and RP1). Power management is handled by ACPI power resources that are shared between the root ports, Thunderbolt controller (NHI) and xHCI controller.
The topology with the power resources (marked with []) looks like:
Host bridge
|
+- RP0 ---\
+- RP1 ---|--+--> [TBT]
+- NHI --/ |
| |
| v
+- xHCI --> [D3C]Here TBT and D3C are the shared ACPI power resources. ACPI _PR3() method returns either TBT or D3C or both.
Say we runtime suspend first the root ports RP0 and RP1, then NHI. Now since the TBT power resource is still on when the root ports are runtime suspended their dev->current_state is set to D3hot. When NHI is runtime suspended TBT is finally turned off but state of the root ports remain to be D3hot.
If the user now runs lspci for instance, the result is all 1's like in the below output (07.0 is the first root port, RP0):
00:07.0 PCI bridge: Intel Corporation Device 8a1d (rev ff) (prog-if ff)
!!! Unknown header type 7f
Kernel driver in use: pcieport
I short the hardware state is not in sync with the software state anymore. The exact same thing happens with the PME polling thread which ends up bringing the root ports back into D0 after they are runtime suspended.
ACPI core already sets the device state to be D3cold when it drops its references to the power resources returned by _PR3 even if these power resources are still physically on (other devices still reference them). However, in PCI core we call acpi_device_get_power() to figure out the power state and that returns the "real" power state based on the state of its power resources.
To make it work with the shared power resources modify acpi_pci_get_power_state() so that it reads the ACPI device power state that was cached by the ACPI core. This makes the PCI device power state match the ACPI device power state regardless of state of the shared power resources that may still be on at this point.
--- drivers/pci/pci-acpi.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/pci/pci-acpi.c +2 -1
@@ -685,7 +685,8 @@ static pci_power_t acpi_pci_get_power_state(struct pci_dev *dev)if (!adev || !acpi_device_power_manageable (adev))return PCI_UNKNOWN ; -if (acpi_device_get_power (adev, &state) || state ==ACPI_STATE_UNKNOWN ) + state = adev->power.state; +if (state ==ACPI_STATE_UNKNOWN )return PCI_UNKNOWN ;return state_conv[state];
Signature · 1 line
2.20.1
Re: [PATCH v2 1/3] PCI / ACPI: Use cached ACPI device state to get PCI device power state
Quoted 3 lines · Mika Westerberg
Intel Ice Lake has an integrated Thunderbolt controller which means that the PCIe topology is extended directly from the two root ports (RP0 and RP1).
A PCIe topology is always extended directly from root ports, regardless of whether a Thunderbolt controller is integrated, so I guess I'm missing the point you're making. It doesn't sound like this is anything specific to Thunderbolt?
Quoted 27 lines
Power management is handled by ACPI power resources that are shared between the root ports, Thunderbolt controller (NHI) and xHCI controller.
The topology with the power resources (marked with []) looks like:
Host bridge | +- RP0 ---\ +- RP1 ---|--+--> [TBT] +- NHI --/ | | | | v +- xHCI --> [D3C]Here TBT and D3C are the shared ACPI power resources. ACPI
_PR3()method returns either TBT or D3C or both.Say we runtime suspend first the root ports RP0 and RP1, then NHI. Now since the TBT power resource is still on when the root ports are runtime suspended their
dev->current_stateis set to D3hot. When NHI is runtime suspended TBT is finally turned off but state of the root ports remain to be D3hot.If the user now runs lspci for instance, the result is all 1's like in the below output (07.0 is the first root port, RP0):
Quoted 8 lines
00:07.0 PCI bridge: Intel Corporation Device 8a1d (rev ff) (prog-if ff)
!!! Unknown header type 7f
Kernel driver in use: pcieportI short the hardware state is not in sync with the software state anymore. The exact same thing happens with the PME polling thread which ends up bringing the root ports back into D0 after they are runtime suspended.
s/I /In /
Re: [PATCH v2 1/3] PCI / ACPI: Use cached ACPI device state to get PCI device power state
Quoted 9 lines · Bjorn Helgaas
Quoted 3 lines · Mika Westerberg
Intel Ice Lake has an integrated Thunderbolt controller which means that the PCIe topology is extended directly from the two root ports (RP0 and RP1).
A PCIe topology is always extended directly from root ports, regardless of whether a Thunderbolt controller is integrated, so I guess I'm missing the point you're making. It doesn't sound like this is anything specific to Thunderbolt?
The point I'm trying to make here is to explain why this is problem now and not with the previous discrete controllers. With the previous there was only a single ACPI power resource for the root port and the Thunderbolt host router was connected to that root port. PCIe hierarchy was extended through downstream ports (not root ports) of that controller (which includes PCIe switch).
Now the thing is part of the SoC so power management is different and causes problems in Linux.
Quoted 31 lines
Quoted 31 lines
Power management is handled by ACPI power resources that are shared between the root ports, Thunderbolt controller (NHI) and xHCI controller.
The topology with the power resources (marked with []) looks like:
Host bridge | +- RP0 ---\ +- RP1 ---|--+--> [TBT] +- NHI --/ | | | | v +- xHCI --> [D3C]Here TBT and D3C are the shared ACPI power resources. ACPI
_PR3()method returns either TBT or D3C or both.Say we runtime suspend first the root ports RP0 and RP1, then NHI. Now since the TBT power resource is still on when the root ports are runtime suspended their
dev->current_stateis set to D3hot. When NHI is runtime suspended TBT is finally turned off but state of the root ports remain to be D3hot.If the user now runs lspci for instance, the result is all 1's like in the below output (07.0 is the first root port, RP0):
00:07.0 PCI bridge: Intel Corporation Device 8a1d (rev ff) (prog-if ff)
!!! Unknown header type 7f
Kernel driver in use: pcieport
Quoted 6 lines
Quoted 4 lines
I short the hardware state is not in sync with the software state anymore. The exact same thing happens with the PME polling thread which ends up bringing the root ports back into D0 after they are runtime suspended.
s/I /In /
Thanks, I'll fix it.
Re: [PATCH v2 1/3] PCI / ACPI: Use cached ACPI device state to get PCI device power state
Quoted 11 lines · Mika Westerberg
Quoted 9 lines · Bjorn Helgaas
Quoted 3 lines · Mika Westerberg
Intel Ice Lake has an integrated Thunderbolt controller which means that the PCIe topology is extended directly from the two root ports (RP0 and RP1).
A PCIe topology is always extended directly from root ports, regardless of whether a Thunderbolt controller is integrated, so I guess I'm missing the point you're making. It doesn't sound like this is anything specific to Thunderbolt?
Quoted 6 lines
The point I'm trying to make here is to explain why this is problem now and not with the previous discrete controllers. With the previous there was only a single ACPI power resource for the root port and the Thunderbolt host router was connected to that root port. PCIe hierarchy was extended through downstream ports (not root ports) of that controller (which includes PCIe switch).
Sounds like you're using "PCIe topology extension" to mean specifically something below a Thunderbolt controller, excluding a subtree below a root port. I don't think the PCI core is aware of that distinction.
Quoted 2 lines
Now the thing is part of the SoC so power management is different and causes problems in Linux.
The SoC is a physical packaging issue that really doesn't enter into the specs directly. I'm trying to get at the logical topology questions in terms of the PCIe and ACPI specs.
I assume we could dream up a non-Thunderbolt topology that would show the same problem?
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
Power management is handled by ACPI power resources that are shared between the root ports, Thunderbolt controller (NHI) and xHCI controller.
The topology with the power resources (marked with []) looks like:
Host bridge | +- RP0 ---\ +- RP1 ---|--+--> [TBT] +- NHI --/ | | | | v +- xHCI --> [D3C]
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Here TBT and D3C are the shared ACPI power resources. ACPI
_PR3()method returns either TBT or D3C or both.
I'm not very familiar with _PR3. I guess this is under an ACPI object representing a PCI device, e.g., \_SB.PCI0.RP0._PR3?
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Say we runtime suspend first the root ports RP0 and RP1, then NHI. Now since the TBT power resource is still on when the root ports are runtime suspended their
dev->current_stateis set to D3hot. When NHI is runtime suspended TBT is finally turned off but state of the root ports remain to be D3hot.
So in this example we might have:
_SB.PCI0.RP0._PR3: TBT
_SB.PCI0.RP1._PR3: TBT
_SB.PCI0.NHI._PR3: TBT
And when Linux figures out that everything depending on TBT is in D3hot, it evaluates TBT._OFF, which puts them all in D3cold? And part of the problem is that they're now in D3cold (where config access doesn't work) but Linux still thinks they're in D3hot (where config access would work)?
I feel like I'm missing something because I don't know how D3C is involved, since you didn't mention suspending xHCI.
And I can't mentally match up the patch with the D3hot/D3cold state change (if indeed that's the problem). If we were updating the path that evaluates _OFF so it changed the power state of all dependent devices, *that* would make a lot of sense to me because it sounds like that's where the physical change happens that makes things out of sync.
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
If the user now runs lspci for instance, the result is all 1's like in the below output (07.0 is the first root port, RP0):
00:07.0 PCI bridge: Intel Corporation Device 8a1d (rev ff) (prog-if ff)
!!! Unknown header type 7f
Kernel driver in use: pcieport
Re: [PATCH v2 1/3] PCI / ACPI: Use cached ACPI device state to get PCI device power state
Quoted 19 lines · Bjorn Helgaas
Quoted 11 lines · Mika Westerberg
Quoted 9 lines · Bjorn Helgaas
Quoted 3 lines · Mika Westerberg
Intel Ice Lake has an integrated Thunderbolt controller which means that the PCIe topology is extended directly from the two root ports (RP0 and RP1).
A PCIe topology is always extended directly from root ports, regardless of whether a Thunderbolt controller is integrated, so I guess I'm missing the point you're making. It doesn't sound like this is anything specific to Thunderbolt?
Quoted 6 lines
The point I'm trying to make here is to explain why this is problem now and not with the previous discrete controllers. With the previous there was only a single ACPI power resource for the root port and the Thunderbolt host router was connected to that root port. PCIe hierarchy was extended through downstream ports (not root ports) of that controller (which includes PCIe switch).
Quoted 4 lines
Sounds like you're using "PCIe topology extension" to mean specifically something below a Thunderbolt controller, excluding a subtree below a root port. I don't think the PCI core is aware of that distinction.
Right it is not.
Quoted 9 lines
Quoted 2 lines
Now the thing is part of the SoC so power management is different and causes problems in Linux.
The SoC is a physical packaging issue that really doesn't enter into the specs directly. I'm trying to get at the logical topology questions in terms of the PCIe and ACPI specs.
I assume we could dream up a non-Thunderbolt topology that would show the same problem?
Yes.
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
Power management is handled by ACPI power resources that are shared between the root ports, Thunderbolt controller (NHI) and xHCI controller.
The topology with the power resources (marked with []) looks like:
Host bridge | +- RP0 ---\ +- RP1 ---|--+--> [TBT] +- NHI --/ | | | | v +- xHCI --> [D3C]
Quoted 5 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Here TBT and D3C are the shared ACPI power resources. ACPI
_PR3()method returns either TBT or D3C or both.I'm not very familiar with
_PR3. I guess this is under an ACPI object representing a PCI device, e.g., \_SB.PCI0.RP0._PR3?
Correct.
Quoted 11 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Say we runtime suspend first the root ports RP0 and RP1, then NHI. Now since the TBT power resource is still on when the root ports are runtime suspended their
dev->current_stateis set to D3hot. When NHI is runtime suspended TBT is finally turned off but state of the root ports remain to be D3hot.So in this example we might have:
_SB.PCI0.RP0._PR3: TBT
_SB.PCI0.RP1._PR3: TBT
_SB.PCI0.NHI._PR3: TBT
and also D3C.
Quoted 5 lines
And when Linux figures out that everything depending on TBT is in D3hot, it evaluates TBT._OFF, which puts them all in D3cold? And part of the problem is that they're now in D3cold (where config access doesn't work) but Linux still thinks they're in D3hot (where config access would work)?
Exactly.
Quoted 2 lines
I feel like I'm missing something because I don't know how D3C is involved, since you didn't mention suspending xHCI.
That's another power resource so we will also have D3C turned off when xHCI gets suspended but I did not want to complicate things too much in the changelog.
Quoted 6 lines
And I can't mentally match up the patch with the D3hot/D3cold state change (if indeed that's the problem). If we were updating the path that evaluates
_OFFso it changed the power state of all dependent devices, *that* would make a lot of sense to me because it sounds like that's where the physical change happens that makes things out of sync.
I did that in the first version [1] but Rafael pointed out that it is racy one way or another [2].
[1] https://www.spinics.net/lists/linux-pci/msg83583.html
[2] https://www.spinics.net/lists/linux-pci/msg83600.html
Re: [PATCH v2 1/3] PCI / ACPI: Use cached ACPI device state to get PCI device power state
Quoted 82 lines · Mika Westerberg
Quoted 19 lines · Bjorn Helgaas
Quoted 11 lines · Mika Westerberg
Quoted 9 lines · Bjorn Helgaas
Quoted 3 lines · Mika Westerberg
Intel Ice Lake has an integrated Thunderbolt controller which means that the PCIe topology is extended directly from the two root ports (RP0 and RP1).
A PCIe topology is always extended directly from root ports, regardless of whether a Thunderbolt controller is integrated, so I guess I'm missing the point you're making. It doesn't sound like this is anything specific to Thunderbolt?
Quoted 6 lines
The point I'm trying to make here is to explain why this is problem now and not with the previous discrete controllers. With the previous there was only a single ACPI power resource for the root port and the Thunderbolt host router was connected to that root port. PCIe hierarchy was extended through downstream ports (not root ports) of that controller (which includes PCIe switch).
Quoted 4 lines
Sounds like you're using "PCIe topology extension" to mean specifically something below a Thunderbolt controller, excluding a subtree below a root port. I don't think the PCI core is aware of that distinction.
Right it is not.
Quoted 9 lines
Quoted 2 lines
Now the thing is part of the SoC so power management is different and causes problems in Linux.
The SoC is a physical packaging issue that really doesn't enter into the specs directly. I'm trying to get at the logical topology questions in terms of the PCIe and ACPI specs.
I assume we could dream up a non-Thunderbolt topology that would show the same problem?
Yes.
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
Power management is handled by ACPI power resources that are shared between the root ports, Thunderbolt controller (NHI) and xHCI controller.
The topology with the power resources (marked with []) looks like:
Host bridge | +- RP0 ---\ +- RP1 ---|--+--> [TBT] +- NHI --/ | | | | v +- xHCI --> [D3C]Quoted 5 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Here TBT and D3C are the shared ACPI power resources. ACPI
_PR3()method returns either TBT or D3C or both.I'm not very familiar with
_PR3. I guess this is under an ACPI object representing a PCI device, e.g., \_SB.PCI0.RP0._PR3?Correct.
Quoted 11 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Say we runtime suspend first the root ports RP0 and RP1, then NHI. Now since the TBT power resource is still on when the root ports are runtime suspended their
dev->current_stateis set to D3hot. When NHI is runtime suspended TBT is finally turned off but state of the root ports remain to be D3hot.So in this example we might have:
_SB.PCI0.RP0._PR3: TBT
_SB.PCI0.RP1._PR3: TBT
_SB.PCI0.NHI._PR3: TBTand also D3C.
Quoted 5 lines
And when Linux figures out that everything depending on TBT is in D3hot, it evaluates TBT._OFF, which puts them all in D3cold? And part of the problem is that they're now in D3cold (where config access doesn't work) but Linux still thinks they're in D3hot (where config access would work)?
Quoted 8 lines
Exactly.
Quoted 2 lines
I feel like I'm missing something because I don't know how D3C is involved, since you didn't mention suspending xHCI.
That's another power resource so we will also have D3C turned off when xHCI gets suspended but I did not want to complicate things too much in the changelog.
If D3C isn't essential to seeing this problem, you could just omit it altogether. I think stripping out anything that's not essential will make it easier to think about the underlying issues.
Quoted 7 lines
Quoted 6 lines
And I can't mentally match up the patch with the D3hot/D3cold state change (if indeed that's the problem). If we were updating the path that evaluates
_OFFso it changed the power state of all dependent devices, *that* would make a lot of sense to me because it sounds like that's where the physical change happens that makes things out of sync.
Quoted 5 lines
I did that in the first version [1] but Rafael pointed out that it is racy one way or another [2].
[1] https://www.spinics.net/lists/linux-pci/msg83583.html
[2] https://www.spinics.net/lists/linux-pci/msg83600.html
Yeah, interesting. It was definitely a much larger patch. I don't know enough to comment on the races. I would wonder whether there's a way to get rid of the caches that become stale, but that's just an idle thought, not a suggestion.
Bjorn
Rafael J. Wysocki
maintainerRe: [PATCH v2 1/3] PCI / ACPI: Use cached ACPI device state to get PCI device power state
Quoted 104 lines · Bjorn Helgaas
Quoted 82 lines · Mika Westerberg
Quoted 19 lines · Bjorn Helgaas
Quoted 11 lines · Mika Westerberg
Quoted 9 lines · Bjorn Helgaas
Quoted 3 lines · Mika Westerberg
Intel Ice Lake has an integrated Thunderbolt controller which means that the PCIe topology is extended directly from the two root ports (RP0 and RP1).
A PCIe topology is always extended directly from root ports, regardless of whether a Thunderbolt controller is integrated, so I guess I'm missing the point you're making. It doesn't sound like this is anything specific to Thunderbolt?
Quoted 6 lines
The point I'm trying to make here is to explain why this is problem now and not with the previous discrete controllers. With the previous there was only a single ACPI power resource for the root port and the Thunderbolt host router was connected to that root port. PCIe hierarchy was extended through downstream ports (not root ports) of that controller (which includes PCIe switch).
Quoted 4 lines
Sounds like you're using "PCIe topology extension" to mean specifically something below a Thunderbolt controller, excluding a subtree below a root port. I don't think the PCI core is aware of that distinction.
Right it is not.
Quoted 9 lines
Quoted 2 lines
Now the thing is part of the SoC so power management is different and causes problems in Linux.
The SoC is a physical packaging issue that really doesn't enter into the specs directly. I'm trying to get at the logical topology questions in terms of the PCIe and ACPI specs.
I assume we could dream up a non-Thunderbolt topology that would show the same problem?
Yes.
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
Power management is handled by ACPI power resources that are shared between the root ports, Thunderbolt controller (NHI) and xHCI controller.
The topology with the power resources (marked with []) looks like:
Host bridge | +- RP0 ---\ +- RP1 ---|--+--> [TBT] +- NHI --/ | | | | v +- xHCI --> [D3C]Quoted 5 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Here TBT and D3C are the shared ACPI power resources. ACPI
_PR3()method returns either TBT or D3C or both.I'm not very familiar with
_PR3. I guess this is under an ACPI object representing a PCI device, e.g., \_SB.PCI0.RP0._PR3?Correct.
Quoted 11 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Say we runtime suspend first the root ports RP0 and RP1, then NHI. Now since the TBT power resource is still on when the root ports are runtime suspended their
dev->current_stateis set to D3hot. When NHI is runtime suspended TBT is finally turned off but state of the root ports remain to be D3hot.So in this example we might have:
_SB.PCI0.RP0._PR3: TBT
_SB.PCI0.RP1._PR3: TBT
_SB.PCI0.NHI._PR3: TBTand also D3C.
Quoted 5 lines
And when Linux figures out that everything depending on TBT is in D3hot, it evaluates TBT._OFF, which puts them all in D3cold? And part of the problem is that they're now in D3cold (where config access doesn't work) but Linux still thinks they're in D3hot (where config access would work)?
Quoted 8 lines
Exactly.
Quoted 2 lines
I feel like I'm missing something because I don't know how D3C is involved, since you didn't mention suspending xHCI.
That's another power resource so we will also have D3C turned off when xHCI gets suspended but I did not want to complicate things too much in the changelog.
If D3C isn't essential to seeing this problem, you could just omit it altogether. I think stripping out anything that's not essential will make it easier to think about the underlying issues.
Quoted 7 lines
Quoted 6 lines
And I can't mentally match up the patch with the D3hot/D3cold state change (if indeed that's the problem). If we were updating the path that evaluates
_OFFso it changed the power state of all dependent devices, *that* would make a lot of sense to me because it sounds like that's where the physical change happens that makes things out of sync.
Quoted 8 lines
Quoted 5 lines
I did that in the first version [1] but Rafael pointed out that it is racy one way or another [2].
[1] https://www.spinics.net/lists/linux-pci/msg83583.html
[2] https://www.spinics.net/lists/linux-pci/msg83600.htmlYeah, interesting. It was definitely a much larger patch. I don't know enough to comment on the races.
Say two power resources are listed by _PR3 for one device (because why not?) and you want to change the device's state to D3cold only if the two power resources are both "off". Then, you need some locking (or equivalent) to synchronize two power resources with each other, so that you can change the devices state when the last of them goes _OFF. Currently, there is no such synchronization between power resources other then the "system_level" value which may not be reliable enough for this type of use.
Or you can say that the device is in D3cold if at least one of the power resources is _OFF, but IMO that may not really be consistent with the view that the "logical" power state of the device should reflect the physical reality accurately.
Quoted 1 line
I would wonder whether there's a way to get rid of the caches that become stale,
I guess what you mean is that the "cached" (or rather "logical" or "expected") power state value may become different from what is returned by acpi_device_get_power() for the device.
The problem here is that acpi_device_get_power() really only should be used for two purposes: (1) To initialize adev->power.state, or to update it via acpi_device_update_power(), and (2) by the "real_power_state" sysfs attribute (of ACPI device objects). The adev->power.state value should be used anywhere else, in principle, so the Mika's patch is correct.
[Note that adev->power.state cannot be updated after calling acpi_device_get_power() to the value returned by it without updating the reference counters of the power resources that are "on" *exactly* because of the problem at hand here.]
Quoted 1 line
but that's just an idle thought, not a suggestion.
After the initialization of the ACPI subsystem, the authoritative source of the ACPI device power state information is adev->power.state. The ACPI subsystem is expected to update this value as needed going forward (including system-wide transitions like resume from S3).
Re: [PATCH v2 1/3] PCI / ACPI: Use cached ACPI device state to get PCI device power state
Quoted 145 lines · Rafael J. Wysocki
Quoted 103 lines · Bjorn Helgaas
Quoted 82 lines · Mika Westerberg
Quoted 19 lines · Bjorn Helgaas
Quoted 11 lines · Mika Westerberg
Quoted 9 lines · Bjorn Helgaas
Quoted 3 lines · Mika Westerberg
Intel Ice Lake has an integrated Thunderbolt controller which means that the PCIe topology is extended directly from the two root ports (RP0 and RP1).
A PCIe topology is always extended directly from root ports, regardless of whether a Thunderbolt controller is integrated, so I guess I'm missing the point you're making. It doesn't sound like this is anything specific to Thunderbolt?
Quoted 6 lines
The point I'm trying to make here is to explain why this is problem now and not with the previous discrete controllers. With the previous there was only a single ACPI power resource for the root port and the Thunderbolt host router was connected to that root port. PCIe hierarchy was extended through downstream ports (not root ports) of that controller (which includes PCIe switch).
Quoted 4 lines
Sounds like you're using "PCIe topology extension" to mean specifically something below a Thunderbolt controller, excluding a subtree below a root port. I don't think the PCI core is aware of that distinction.
Right it is not.
Quoted 9 lines
Quoted 2 lines
Now the thing is part of the SoC so power management is different and causes problems in Linux.
The SoC is a physical packaging issue that really doesn't enter into the specs directly. I'm trying to get at the logical topology questions in terms of the PCIe and ACPI specs.
I assume we could dream up a non-Thunderbolt topology that would show the same problem?
Yes.
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
Power management is handled by ACPI power resources that are shared between the root ports, Thunderbolt controller (NHI) and xHCI controller.
The topology with the power resources (marked with []) looks like:
Host bridge | +- RP0 ---\ +- RP1 ---|--+--> [TBT] +- NHI --/ | | | | v +- xHCI --> [D3C]Quoted 5 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Here TBT and D3C are the shared ACPI power resources. ACPI
_PR3()method returns either TBT or D3C or both.I'm not very familiar with
_PR3. I guess this is under an ACPI object representing a PCI device, e.g., \_SB.PCI0.RP0._PR3?Correct.
Quoted 11 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Say we runtime suspend first the root ports RP0 and RP1, then NHI. Now since the TBT power resource is still on when the root ports are runtime suspended their
dev->current_stateis set to D3hot. When NHI is runtime suspended TBT is finally turned off but state of the root ports remain to be D3hot.So in this example we might have:
_SB.PCI0.RP0._PR3: TBT
_SB.PCI0.RP1._PR3: TBT
_SB.PCI0.NHI._PR3: TBTand also D3C.
Quoted 5 lines
And when Linux figures out that everything depending on TBT is in D3hot, it evaluates TBT._OFF, which puts them all in D3cold? And part of the problem is that they're now in D3cold (where config access doesn't work) but Linux still thinks they're in D3hot (where config access would work)?
Quoted 8 lines
Exactly.
Quoted 2 lines
I feel like I'm missing something because I don't know how D3C is involved, since you didn't mention suspending xHCI.
That's another power resource so we will also have D3C turned off when xHCI gets suspended but I did not want to complicate things too much in the changelog.
If D3C isn't essential to seeing this problem, you could just omit it altogether. I think stripping out anything that's not essential will make it easier to think about the underlying issues.
Quoted 7 lines
Quoted 6 lines
And I can't mentally match up the patch with the D3hot/D3cold state change (if indeed that's the problem). If we were updating the path that evaluates
_OFFso it changed the power state of all dependent devices, *that* would make a lot of sense to me because it sounds like that's where the physical change happens that makes things out of sync.Quoted 8 lines
Quoted 5 lines
I did that in the first version [1] but Rafael pointed out that it is racy one way or another [2].
[1] https://www.spinics.net/lists/linux-pci/msg83583.html
[2] https://www.spinics.net/lists/linux-pci/msg83600.htmlYeah, interesting. It was definitely a much larger patch. I don't know enough to comment on the races.
Say two power resources are listed by
_PR3for one device (because why not?) and you want to change the device's state to D3cold only if the two power resources are both "off". Then, you need some locking (or equivalent) to synchronize two power resources with each other, so that you can change the devices state when the last of them goes_OFF. Currently, there is no such synchronization between power resources other then the "system_level" value which may not be reliable enough for this type of use.Or you can say that the device is in D3cold if at least one of the power resources is
_OFF, but IMO that may not really be consistent with the view that the "logical" power state of the device should reflect the physical reality accurately.Quoted 1 line
I would wonder whether there's a way to get rid of the caches that become stale,
I guess what you mean is that the "cached" (or rather "logical" or "expected") power state value may become different from what is returned by
acpi_device_get_power()for the device.The problem here is that
acpi_device_get_power()really only should be used for two purposes: (1) To initializeadev->power.state, or to update it viaacpi_device_update_power(), and (2) by the "real_power_state" sysfs attribute (of ACPI device objects). Theadev->power.statevalue should be used anywhere else, in principle, so the Mika's patch is correct.[Note that
adev->power.statecannot be updated after callingacpi_device_get_power()to the value returned by it without updating the reference counters of the power resources that are "on" *exactly* because of the problem at hand here.]
Quoted 7 lines
Quoted 1 line
but that's just an idle thought, not a suggestion.
After the initialization of the ACPI subsystem, the authoritative source of the ACPI device power state information is
adev->power.state. The ACPI subsystem is expected to update this value as needed going forward (including system-wide transitions like resume from S3).
Thanks, this is all very helpful! Do you by any chance add lore.kernel.org links to commit logs when applying patches? This is a case where I think the discussion could be useful in the future.
Re: [PATCH v2 1/3] PCI / ACPI: Use cached ACPI device state to get PCI device power state
Quoted 154 lines · CEST Bjorn Helgaas
Quoted 145 lines · Rafael J. Wysocki
Quoted 103 lines · Bjorn Helgaas
Quoted 82 lines · Mika Westerberg
Quoted 19 lines · Bjorn Helgaas
Quoted 11 lines · Mika Westerberg
Quoted 9 lines · Bjorn Helgaas
Quoted 3 lines · Mika Westerberg
Intel Ice Lake has an integrated Thunderbolt controller which means that the PCIe topology is extended directly from the two root ports (RP0 and RP1).
A PCIe topology is always extended directly from root ports, regardless of whether a Thunderbolt controller is integrated, so I guess I'm missing the point you're making. It doesn't sound like this is anything specific to Thunderbolt?
Quoted 6 lines
The point I'm trying to make here is to explain why this is problem now and not with the previous discrete controllers. With the previous there was only a single ACPI power resource for the root port and the Thunderbolt host router was connected to that root port. PCIe hierarchy was extended through downstream ports (not root ports) of that controller (which includes PCIe switch).
Quoted 4 lines
Sounds like you're using "PCIe topology extension" to mean specifically something below a Thunderbolt controller, excluding a subtree below a root port. I don't think the PCI core is aware of that distinction.
Right it is not.
Quoted 9 lines
Quoted 2 lines
Now the thing is part of the SoC so power management is different and causes problems in Linux.
The SoC is a physical packaging issue that really doesn't enter into the specs directly. I'm trying to get at the logical topology questions in terms of the PCIe and ACPI specs.
I assume we could dream up a non-Thunderbolt topology that would show the same problem?
Yes.
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
Power management is handled by ACPI power resources that are shared between the root ports, Thunderbolt controller (NHI) and xHCI controller.
The topology with the power resources (marked with []) looks like:
Host bridge | +- RP0 ---\ +- RP1 ---|--+--> [TBT] +- NHI --/ | | | | v +- xHCI --> [D3C]Quoted 5 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Here TBT and D3C are the shared ACPI power resources. ACPI
_PR3()method returns either TBT or D3C or both.I'm not very familiar with
_PR3. I guess this is under an ACPI object representing a PCI device, e.g., \_SB.PCI0.RP0._PR3?Correct.
Quoted 11 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Say we runtime suspend first the root ports RP0 and RP1, then NHI. Now since the TBT power resource is still on when the root ports are runtime suspended their
dev->current_stateis set to D3hot. When NHI is runtime suspended TBT is finally turned off but state of the root ports remain to be D3hot.So in this example we might have:
_SB.PCI0.RP0._PR3: TBT
_SB.PCI0.RP1._PR3: TBT
_SB.PCI0.NHI._PR3: TBTand also D3C.
Quoted 5 lines
And when Linux figures out that everything depending on TBT is in D3hot, it evaluates TBT._OFF, which puts them all in D3cold? And part of the problem is that they're now in D3cold (where config access doesn't work) but Linux still thinks they're in D3hot (where config access would work)?
Quoted 8 lines
Exactly.
Quoted 2 lines
I feel like I'm missing something because I don't know how D3C is involved, since you didn't mention suspending xHCI.
That's another power resource so we will also have D3C turned off when xHCI gets suspended but I did not want to complicate things too much in the changelog.
If D3C isn't essential to seeing this problem, you could just omit it altogether. I think stripping out anything that's not essential will make it easier to think about the underlying issues.
Quoted 7 lines
Quoted 6 lines
And I can't mentally match up the patch with the D3hot/D3cold state change (if indeed that's the problem). If we were updating the path that evaluates
_OFFso it changed the power state of all dependent devices, *that* would make a lot of sense to me because it sounds like that's where the physical change happens that makes things out of sync.Quoted 8 lines
Quoted 5 lines
I did that in the first version [1] but Rafael pointed out that it is racy one way or another [2].
[1] https://www.spinics.net/lists/linux-pci/msg83583.html
[2] https://www.spinics.net/lists/linux-pci/msg83600.htmlYeah, interesting. It was definitely a much larger patch. I don't know enough to comment on the races.
Say two power resources are listed by
_PR3for one device (because why not?) and you want to change the device's state to D3cold only if the two power resources are both "off". Then, you need some locking (or equivalent) to synchronize two power resources with each other, so that you can change the devices state when the last of them goes_OFF. Currently, there is no such synchronization between power resources other then the "system_level" value which may not be reliable enough for this type of use.Or you can say that the device is in D3cold if at least one of the power resources is
_OFF, but IMO that may not really be consistent with the view that the "logical" power state of the device should reflect the physical reality accurately.Quoted 1 line
I would wonder whether there's a way to get rid of the caches that become stale,
I guess what you mean is that the "cached" (or rather "logical" or "expected") power state value may become different from what is returned by
acpi_device_get_power()for the device.The problem here is that
acpi_device_get_power()really only should be used for two purposes: (1) To initializeadev->power.state, or to update it viaacpi_device_update_power(), and (2) by the "real_power_state" sysfs attribute (of ACPI device objects). Theadev->power.statevalue should be used anywhere else, in principle, so the Mika's patch is correct.[Note that
adev->power.statecannot be updated after callingacpi_device_get_power()to the value returned by it without updating the reference counters of the power resources that are "on" *exactly* because of the problem at hand here.]Quoted 7 lines
Quoted 1 line
but that's just an idle thought, not a suggestion.
After the initialization of the ACPI subsystem, the authoritative source of the ACPI device power state information is
adev->power.state. The ACPI subsystem is expected to update this value as needed going forward (including system-wide transitions like resume from S3).
Quoted 5 lines
Thanks, this is all very helpful! Do you by any chance add lore.kernel.org links to commit logs when applying patches? This is a case where I think the discussion could be useful in the future.
Agreed, and thanks for the URL.
I guess Mika can add this tag to the patch changelog.
Re: [PATCH v2 1/3] PCI / ACPI: Use cached ACPI device state to get PCI device power state
Quoted 9 lines · Rafael J. Wysocki
Quoted 5 lines
Thanks, this is all very helpful! Do you by any chance add lore.kernel.org links to commit logs when applying patches? This is a case where I think the discussion could be useful in the future.
Agreed, and thanks for the URL.
I guess Mika can add this tag to the patch changelog.
Sure I'll add it.
Re: [PATCH v2 1/3] PCI / ACPI: Use cached ACPI device state to get PCI device power state
In reply to Rafael J. WysockiQuoted 134 lines · CEST Rafael J. Wysocki
Quoted 104 lines · Bjorn Helgaas
Quoted 82 lines · Mika Westerberg
Quoted 19 lines · Bjorn Helgaas
Quoted 11 lines · Mika Westerberg
Quoted 9 lines · Bjorn Helgaas
Quoted 3 lines · Mika Westerberg
Intel Ice Lake has an integrated Thunderbolt controller which means that the PCIe topology is extended directly from the two root ports (RP0 and RP1).
A PCIe topology is always extended directly from root ports, regardless of whether a Thunderbolt controller is integrated, so I guess I'm missing the point you're making. It doesn't sound like this is anything specific to Thunderbolt?
Quoted 6 lines
The point I'm trying to make here is to explain why this is problem now and not with the previous discrete controllers. With the previous there was only a single ACPI power resource for the root port and the Thunderbolt host router was connected to that root port. PCIe hierarchy was extended through downstream ports (not root ports) of that controller (which includes PCIe switch).
Quoted 4 lines
Sounds like you're using "PCIe topology extension" to mean specifically something below a Thunderbolt controller, excluding a subtree below a root port. I don't think the PCI core is aware of that distinction.
Right it is not.
Quoted 9 lines
Quoted 2 lines
Now the thing is part of the SoC so power management is different and causes problems in Linux.
The SoC is a physical packaging issue that really doesn't enter into the specs directly. I'm trying to get at the logical topology questions in terms of the PCIe and ACPI specs.
I assume we could dream up a non-Thunderbolt topology that would show the same problem?
Yes.
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
Power management is handled by ACPI power resources that are shared between the root ports, Thunderbolt controller (NHI) and xHCI controller.
The topology with the power resources (marked with []) looks like:
Host bridge | +- RP0 ---\ +- RP1 ---|--+--> [TBT] +- NHI --/ | | | | v +- xHCI --> [D3C]Quoted 5 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Here TBT and D3C are the shared ACPI power resources. ACPI
_PR3()method returns either TBT or D3C or both.I'm not very familiar with
_PR3. I guess this is under an ACPI object representing a PCI device, e.g., \_SB.PCI0.RP0._PR3?Correct.
Quoted 11 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Say we runtime suspend first the root ports RP0 and RP1, then NHI. Now since the TBT power resource is still on when the root ports are runtime suspended their
dev->current_stateis set to D3hot. When NHI is runtime suspended TBT is finally turned off but state of the root ports remain to be D3hot.So in this example we might have:
_SB.PCI0.RP0._PR3: TBT
_SB.PCI0.RP1._PR3: TBT
_SB.PCI0.NHI._PR3: TBTand also D3C.
Quoted 5 lines
And when Linux figures out that everything depending on TBT is in D3hot, it evaluates TBT._OFF, which puts them all in D3cold? And part of the problem is that they're now in D3cold (where config access doesn't work) but Linux still thinks they're in D3hot (where config access would work)?
Quoted 8 lines
Exactly.
Quoted 2 lines
I feel like I'm missing something because I don't know how D3C is involved, since you didn't mention suspending xHCI.
That's another power resource so we will also have D3C turned off when xHCI gets suspended but I did not want to complicate things too much in the changelog.
If D3C isn't essential to seeing this problem, you could just omit it altogether. I think stripping out anything that's not essential will make it easier to think about the underlying issues.
Quoted 7 lines
Quoted 6 lines
And I can't mentally match up the patch with the D3hot/D3cold state change (if indeed that's the problem). If we were updating the path that evaluates
_OFFso it changed the power state of all dependent devices, *that* would make a lot of sense to me because it sounds like that's where the physical change happens that makes things out of sync.Quoted 8 lines
Quoted 5 lines
I did that in the first version [1] but Rafael pointed out that it is racy one way or another [2].
[1] https://www.spinics.net/lists/linux-pci/msg83583.html
[2] https://www.spinics.net/lists/linux-pci/msg83600.htmlYeah, interesting. It was definitely a much larger patch. I don't know enough to comment on the races.
Say two power resources are listed by
_PR3for one device (because why not?) and you want to change the device's state to D3cold only if the two power resources are both "off". Then, you need some locking (or equivalent) to synchronize two power resources with each other, so that you can change the devices state when the last of them goes_OFF. Currently, there is no such synchronization between power resources other then the "system_level" value which may not be reliable enough for this type of use.Or you can say that the device is in D3cold if at least one of the power resources is
_OFF, but IMO that may not really be consistent with the view that the "logical" power state of the device should reflect the physical reality accurately.Quoted 1 line
I would wonder whether there's a way to get rid of the caches that become stale,
I guess what you mean is that the "cached" (or rather "logical" or "expected") power state value may become different from what is returned by
acpi_device_get_power()for the device.
Quoted 6 lines
The problem here is that
acpi_device_get_power()really only should be used for two purposes: (1) To initializeadev->power.state, or to update it viaacpi_device_update_power(), and (2) by the "real_power_state" sysfs attribute (of ACPI device objects). Theadev->power.statevalue should be used anywhere else, in principle, so the Mika's patch is correct.
Well, it is an improvement, but it is not sufficient.
Quoted 4 lines
[Note that
adev->power.statecannot be updated after callingacpi_device_get_power()to the value returned by it without updating the reference counters of the power resources that are "on" *exactly* because of the problem at hand here.]
That is obviously correct, but ->
Quoted 7 lines
Quoted 1 line
but that's just an idle thought, not a suggestion.
After the initialization of the ACPI subsystem, the authoritative source of the ACPI device power state information is
adev->power.state. The ACPI subsystem is expected to update this value as needed going forward (including system-wide transitions like resume from S3).
-> the "resume from S3 or hibernation" case needs special handling, because in that case the device power state need not reflect the information the ACPI subsystem has. That only matters if adev->power.state is ACPI_STATE_D0 and the device is actually *not* in D0, because in that case acpi_device_set_power() will not work. So that case is not covered currently (it should be rare in practice, though, if it happens at all), so something like the patch below (untested) may be needed in addition to the Mika's patch.
Still, there is also the "power state not matching" case in pci_pm_complete() that's need to be covered and the non-PCI ACPI PM has a similar issue in theory, so I need to think about this a bit more.
--- drivers/pci/pci-acpi.c | 14 +++++++++++++- drivers/pci/pci-mid.c | 3 ++- drivers/pci/pci.c | 9 +++++---- drivers/pci/pci.h | 2 +- 4 files changed, 21 insertions(+), 7 deletions(-)
drivers/pci/pci-acpi.c +13 -1
@@ -632,7 +632,8 @@ static bool acpi_pci_power_manageable(streturn adev ?acpi_device_power_manageable (adev) : false; } -static int acpi_pci_set_power_state (struct pci_dev *dev,pci_power_t state) +static int acpi_pci_set_power_state (struct pci_dev *dev,pci_power_t state, +bool force) {struct acpi_device *adev =ACPI_COMPANION (&dev->dev);static const u8 state_conv[] = {
@@ -657,6 +658,17 @@ static int acpi_pci_set_power_state(stru }/* Fall through */ case PCI_D0 : +if (force) { +int acpi_state; + + error =acpi_device_update_power (adev, &acpi_state); +if (error) +return error; + +if (acpi_state ==ACPI_STATE_D0 ) +return 0 ; + } +/* fall through */ case PCI_D1 :case PCI_D2 :case PCI_D3hot:
drivers/pci/pci-mid.c +2 -1
@@ -21,7 +21,8 @@ static bool mid_pci_power_manageable(strreturn true; } -static int mid_pci_set_power_state (struct pci_dev *pdev,pci_power_t state) +static int mid_pci_set_power_state (struct pci_dev *pdev,pci_power_t state, +bool not_used) {return intel_mid_pci_set_power_state (pdev, state); }
drivers/pci/pci.c +5 -4
@@ -767,9 +767,10 @@ static inline bool platform_pci_power_ma }static inline int platform_pci_set_power_state (struct pci_dev *dev, -pci_power_t t) +pci_power_t state,bool force) { -return pci_platform_pm ? pci_platform_pm->set_state (dev, t) : -ENOSYS ; +return pci_platform_pm ? + pci_platform_pm->set_state (dev, state, force) : -ENOSYS ; }static inline pci_power_t platform_pci_get_power_state (struct pci_dev *dev)
@@ -944,7 +945,7 @@ void pci_update_current_state(struct pcivoid pci_power_up (struct pci_dev *dev) {if (platform_pci_power_manageable (dev)) -platform_pci_set_power_state (dev,PCI_D0 ); +platform_pci_set_power_state (dev,PCI_D0 , true);pci_raw_set_power_state (dev,PCI_D0 );pci_update_current_state (dev,PCI_D0 );
@@ -960,7 +961,7 @@ static int pci_platform_power_transitionint error;if (platform_pci_power_manageable (dev)) { - error =platform_pci_set_power_state (dev, state); + error =platform_pci_set_power_state (dev, state, false);if (!error)pci_update_current_state (dev, state); }else
drivers/pci/pci.h +1 -1
@@ -67,7 +67,7 @@ int pci_bus_error_reset(struct pci_dev *struct pci_platform_pm_ops {bool (*bridge_d3)(struct pci_dev *dev);bool (*is_manageable)(struct pci_dev *dev); -int (*set_state)(struct pci_dev *dev,pci_power_t state); +int (*set_state)(struct pci_dev *dev,pci_power_t state,bool force);pci_power_t (*get_state)(struct pci_dev *dev);pci_power_t (*choose_state)(struct pci_dev *dev);int (*set_wakeup)(struct pci_dev *dev,bool enable);
Re: [PATCH v2 1/3] PCI / ACPI: Use cached ACPI device state to get PCI device power state
Quoted 24 lines · Rafael J. Wysocki
Quoted 6 lines
The problem here is that
acpi_device_get_power()really only should be used for two purposes: (1) To initializeadev->power.state, or to update it viaacpi_device_update_power(), and (2) by the "real_power_state" sysfs attribute (of ACPI device objects). Theadev->power.statevalue should be used anywhere else, in principle, so the Mika's patch is correct.Well, it is an improvement, but it is not sufficient.
Quoted 4 lines
[Note that
adev->power.statecannot be updated after callingacpi_device_get_power()to the value returned by it without updating the reference counters of the power resources that are "on" *exactly* because of the problem at hand here.]That is obviously correct, but ->
Quoted 7 lines
Quoted 1 line
but that's just an idle thought, not a suggestion.
After the initialization of the ACPI subsystem, the authoritative source of the ACPI device power state information is
adev->power.state. The ACPI subsystem is expected to update this value as needed going forward (including system-wide transitions like resume from S3).
Quoted 5 lines
-> the "resume from S3 or hibernation" case needs special handling, because in that case the device power state need not reflect the information the ACPI subsystem has. That only matters if
adev->power.stateisACPI_STATE_D0and the device is actually *not* in D0, because in that caseacpi_device_set_power()will not work.
I guess you are talking about the special-cased devices that we leave in D0 when system suspend (via firmware) is entered?
Quoted 3 lines
So that case is not covered currently (it should be rare in practice, though, if it happens at all), so something like the patch below (untested) may be needed in addition to the Mika's patch.
Looks good to me.
Quoted 3 lines
Still, there is also the "power state not matching" case in
pci_pm_complete()that's need to be covered and the non-PCI ACPI PM has a similar issue in theory, so I need to think about this a bit more.
Do you want me to hold off sending an updated version of the patch series while we figure this one out or is it fine if I send it out now and we can add further details on top?
Rafael J. Wysocki
maintainerRe: [PATCH v2 1/3] PCI / ACPI: Use cached ACPI device state to get PCI device power state
Quoted 32 lines · Mika Westerberg
Quoted 24 lines · Rafael J. Wysocki
Quoted 6 lines
The problem here is that
acpi_device_get_power()really only should be used for two purposes: (1) To initializeadev->power.state, or to update it viaacpi_device_update_power(), and (2) by the "real_power_state" sysfs attribute (of ACPI device objects). Theadev->power.statevalue should be used anywhere else, in principle, so the Mika's patch is correct.Well, it is an improvement, but it is not sufficient.
Quoted 4 lines
[Note that
adev->power.statecannot be updated after callingacpi_device_get_power()to the value returned by it without updating the reference counters of the power resources that are "on" *exactly* because of the problem at hand here.]That is obviously correct, but ->
Quoted 7 lines
Quoted 1 line
but that's just an idle thought, not a suggestion.
After the initialization of the ACPI subsystem, the authoritative source of the ACPI device power state information is
adev->power.state. The ACPI subsystem is expected to update this value as needed going forward (including system-wide transitions like resume from S3).Quoted 5 lines
-> the "resume from S3 or hibernation" case needs special handling, because in that case the device power state need not reflect the information the ACPI subsystem has. That only matters if
adev->power.stateisACPI_STATE_D0and the device is actually *not* in D0, because in that caseacpi_device_set_power()will not work.
Quoted 8 lines
I guess you are talking about the special-cased devices that we leave in D0 when system suspend (via firmware) is entered?
Quoted 3 lines
So that case is not covered currently (it should be rare in practice, though, if it happens at all), so something like the patch below (untested) may be needed in addition to the Mika's patch.
Looks good to me.
I actually decided to address this issue in acpi_device_set_power() as it may affect devices beyond PCI in principle. I will send a patch for that shortly.
Quoted 7 lines
Quoted 3 lines
Still, there is also the "power state not matching" case in
pci_pm_complete()that's need to be covered and the non-PCI ACPI PM has a similar issue in theory, so I need to think about this a bit more.Do you want me to hold off sending an updated version of the patch series while we figure this one out or is it fine if I send it out now and we can add further details on top?
It is independent of the other fix, so it can be sent now just fine IMO.
Re: [PATCH v2 1/3] PCI / ACPI: Use cached ACPI device state to get PCI device power state
Quoted 41 lines · Rafael J. Wysocki
Quoted 39 lines · Mika Westerberg
Quoted 24 lines · Rafael J. Wysocki
Quoted 6 lines
The problem here is that
acpi_device_get_power()really only should be used for two purposes: (1) To initializeadev->power.state, or to update it viaacpi_device_update_power(), and (2) by the "real_power_state" sysfs attribute (of ACPI device objects). Theadev->power.statevalue should be used anywhere else, in principle, so the Mika's patch is correct.Well, it is an improvement, but it is not sufficient.
Quoted 4 lines
[Note that
adev->power.statecannot be updated after callingacpi_device_get_power()to the value returned by it without updating the reference counters of the power resources that are "on" *exactly* because of the problem at hand here.]That is obviously correct, but ->
Quoted 7 lines
Quoted 1 line
but that's just an idle thought, not a suggestion.
After the initialization of the ACPI subsystem, the authoritative source of the ACPI device power state information is
adev->power.state. The ACPI subsystem is expected to update this value as needed going forward (including system-wide transitions like resume from S3).Quoted 5 lines
-> the "resume from S3 or hibernation" case needs special handling, because in that case the device power state need not reflect the information the ACPI subsystem has. That only matters if
adev->power.stateisACPI_STATE_D0and the device is actually *not* in D0, because in that caseacpi_device_set_power()will not work.I guess you are talking about the special-cased devices that we leave in D0 when system suspend (via firmware) is entered?
Quoted 3 lines
So that case is not covered currently (it should be rare in practice, though, if it happens at all), so something like the patch below (untested) may be needed in addition to the Mika's patch.
Quoted 5 lines
Quoted 1 line
Looks good to me.
I actually decided to address this issue in
acpi_device_set_power()as it may affect devices beyond PCI in principle. I will send a patch for that shortly.
Thanks!
Quoted 9 lines
Quoted 7 lines
Quoted 3 lines
Still, there is also the "power state not matching" case in
pci_pm_complete()that's need to be covered and the non-PCI ACPI PM has a similar issue in theory, so I need to think about this a bit more.Do you want me to hold off sending an updated version of the patch series while we figure this one out or is it fine if I send it out now and we can add further details on top?
It is independent of the other fix, so it can be sent now just fine IMO.
OK, I'll send it out in a minute.
Rafael J. Wysocki
maintainerRe: [PATCH v2 1/3] PCI / ACPI: Use cached ACPI device state to get PCI device power state
In reply to Mika WesterbergQuoted 1 line · Mika Westerberg
Actually, to start with, you can say that the ACPI power state returned by acpi_device_get_power() may depend on the configuration of ACPI power resources in the system which may change at any time after acpi_device_get_power() has returned, unless the reference counters of the ACPI power resources in question are set to prevent that from happening. Thus it is invalid to use acpi_device_get_power() in acpi_pci_get_power_state() the way it is done now and the value of the power.state field in the corresponding struct acpi_device object (which reflects the ACPI power resources reference counting, among other things) should be used instead.
Then you can describe the particular issue below as an example.
IMO that would explain the rationale better here.
Quoted 73 lines
Intel Ice Lake has an integrated Thunderbolt controller which means that the PCIe topology is extended directly from the two root ports (RP0 and RP1). Power management is handled by ACPI power resources that are shared between the root ports, Thunderbolt controller (NHI) and xHCI controller.
The topology with the power resources (marked with []) looks like:
Host bridge | +- RP0 ---\ +- RP1 ---|--+--> [TBT] +- NHI --/ | | | | v +- xHCI --> [D3C]Here TBT and D3C are the shared ACPI power resources. ACPI
_PR3()method returns either TBT or D3C or both.Say we runtime suspend first the root ports RP0 and RP1, then NHI. Now since the TBT power resource is still on when the root ports are runtime suspended their
dev->current_stateis set to D3hot. When NHI is runtime suspended TBT is finally turned off but state of the root ports remain to be D3hot.If the user now runs lspci for instance, the result is all 1's like in the below output (07.0 is the first root port, RP0):
00:07.0 PCI bridge: Intel Corporation Device 8a1d (rev ff) (prog-if ff)
!!! Unknown header type 7f
Kernel driver in use: pcieportI short the hardware state is not in sync with the software state anymore. The exact same thing happens with the PME polling thread which ends up bringing the root ports back into D0 after they are runtime suspended.
ACPI core already sets the device state to be D3cold when it drops its references to the power resources returned by
_PR3even if these power resources are still physically on (other devices still reference them). However, in PCI core we callacpi_device_get_power()to figure out the power state and that returns the "real" power state based on the state of its power resources.To make it work with the shared power resources modify
acpi_pci_get_power_state()so that it reads the ACPI device power state that was cached by the ACPI core. This makes the PCI device power state match the ACPI device power state regardless of state of the shared power resources that may still be on at this point.Signed-off-by Mika Westerberg <[email protected]>--- drivers/pci/pci-acpi.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)drivers/pci/pci-acpi.c +2
-2@@ -685,7 +685,8 @@ static pci_power_t acpi_pci_get_power_state(struct pci_dev *dev)if (!adev || !acpi_device_power_manageable (adev))return PCI_UNKNOWN ; -if (acpi_device_get_power (adev, &state) || state ==ACPI_STATE_UNKNOWN ) + state = adev->power.state; +if (state ==ACPI_STATE_UNKNOWN )return PCI_UNKNOWN ;return state_conv[state]; --2.20.1
Re: [PATCH v2 1/3] PCI / ACPI: Use cached ACPI device state to get PCI device power state
Quoted 15 lines · Rafael J. Wysocki
Quoted 1 line · Mika Westerberg
Actually, to start with, you can say that the ACPI power state returned by
acpi_device_get_power()may depend on the configuration of ACPI power resources in the system which may change at any time afteracpi_device_get_power()has returned, unless the reference counters of the ACPI power resources in question are set to prevent that from happening. Thus it is invalid to useacpi_device_get_power()inacpi_pci_get_power_state()the way it is done now and the value of the power.state field in the corresponding structacpi_deviceobject (which reflects the ACPI power resources reference counting, among other things) should be used instead.
Quoted 3 lines
Then you can describe the particular issue below as an example.
IMO that would explain the rationale better here.
Thanks! I'll update the changelog accordingly.
PATCH v2 2/3 ACPI / PM: Introduce concept of a _PR0 dependent device
In reply to Mika WesterbergIf there are shared power resources between otherwise unrelated devices turning them on causes the other devices sharing them to be powered up as well. In case of PCI devices go into D0uninitialized state meaning that if they were configured to trigger wake that configuration is lost at this point.
For this reason introduce a concept of "_PR0 dependent device" that can be added to any ACPI device that has power resources. The dependent device will be included in a list of dependent devices for all power resources returned by the ACPI device's _PR0 (assuming it has one). Whenever a power resource having dependent devices is turned physically on (its _ON method is called) we runtime resume all of them to allow their driver or in case of PCI the PCI core to re-initialize the device and its wake configuration.
This adds two functions that can be used to add and remove these dependent devices. Note the dependent device does not necessary need share power resources so this functionality can be used to add "software dependencies" as well if needed.
--- drivers/acpi/power.c | 139 ++++++++++++++++++++++++++++++++++++++++ include/acpi/acpi_bus.h | 4 ++ 2 files changed, 143 insertions(+)
drivers/acpi/power.c +139 -0
@@ -42,6 +42,11 @@ ACPI_MODULE_NAME("power");#define ACPI_POWER_RESOURCE_STATE_ON 0x01#define ACPI_POWER_RESOURCE_STATE_UNKNOWN 0xFF +struct acpi_power_dependent_device { +struct device *dev; +struct list_head node; +}; +struct acpi_power_resource {struct acpi_device device;struct list_head list_node;
@@ -51,6 +56,7 @@ struct acpi_power_resource {unsigned int ref_count;bool wakeup_enabled;struct mutex resource_lock; +struct list_head dependents; };struct acpi_power_resource_entry {
@@ -232,8 +238,125 @@ static int acpi_power_get_list_state(struct list_head *list, int *state)return 0 ; } +static int +acpi_power_resource_add_dependent (struct acpi_power_resource *resource, +struct device *dev) +{ +struct acpi_power_dependent_device *dep; +int ret =0 ; + +mutex_lock (&resource->resource_lock); +list_for_each_entry (dep, &resource->dependents, node) { +/* Only add it once */ +if (dep->dev == dev) + goto unlock; + } + + dep =kzalloc (sizeof (*dep),GFP_KERNEL ); +if (!dep) { + ret = -ENOMEM ; + goto unlock; + } + + dep->dev = dev; +list_add_tail (&dep->node, &resource->dependents); +dev_dbg (dev,"added power dependency to [%s]\n" , resource->name); + +unlock: +mutex_unlock (&resource->resource_lock); +return ret; +} + +static void +acpi_power_resource_remove_dependent (struct acpi_power_resource *resource, +struct device *dev) +{ +struct acpi_power_dependent_device *dep; + +mutex_lock (&resource->resource_lock); +list_for_each_entry (dep, &resource->dependents, node) { +if (dep->dev == dev) { +list_del (&dep->node); +kfree (dep); +dev_dbg (dev,"removed power dependency to [%s]\n" , + resource->name); +break ; + } + } +mutex_unlock (&resource->resource_lock); +} + +/** +* acpi_device_power_add_dependent - Add dependent device of this ACPI device +* @adev: ACPI device pointer +* @dev: Dependent device +* +* If @adev has non-empty _PR0 the @dev is added as dependent device to all +* power resources returned by it. This means that whenever these power +* resources are turned _ON the dependent devices get runtime resumed. This +* is needed for devices such as PCI to allow its driver to re-initialize +* it after it went to D0uninitialized. +* +* If @adev does not have _PR0 this does nothing. +* +* Returns %0 in case of success and negative errno otherwise. +*/ +int acpi_device_power_add_dependent (struct acpi_device *adev, +struct device *dev) +{ +struct acpi_power_resource_entry *entry; +struct list_head *resources; +int ret; + +if (!adev->power.flags.power_resources) +return 0 ; +if (!adev->power.states[ACPI_STATE_D0 ].flags.valid) +return 0 ; + + resources = &adev->power.states[ACPI_STATE_D0 ].resources; +list_for_each_entry (entry, resources, node) { + ret =acpi_power_resource_add_dependent (entry->resource, dev); +if (ret) + goto err; + } + +return 0 ; + +err: +list_for_each_entry (entry, resources, node) +acpi_power_resource_remove_dependent (entry->resource, dev); + +return ret; +} + +/** +* acpi_device_power_remove_dependent - Remove dependent device +* @adev: ACPI device pointer +* @dev: Dependent device +* +* Does the opposite of acpi_device_power_add_dependent() and removes the +* dependent device if it is found. Can be called to @adev that does not +* have _PR0 as well. +*/ +void acpi_device_power_remove_dependent (struct acpi_device *adev, +struct device *dev) +{ +struct acpi_power_resource_entry *entry; +struct list_head *resources; + +if (!adev->power.flags.power_resources) +return ; +if (!adev->power.states[ACPI_STATE_D0 ].flags.valid) +return ; + + resources = &adev->power.states[ACPI_STATE_D0 ].resources; +list_for_each_entry_reverse (entry, resources, node) +acpi_power_resource_remove_dependent (entry->resource, dev); +} +static int __acpi_power_on (struct acpi_power_resource *resource) { +struct acpi_power_dependent_device *dep;acpi_status status =AE_OK ; status =acpi_evaluate_object (resource->device.handle,"_ON" ,NULL ,NULL );
@@ -243,6 +366,21 @@ static int __acpi_power_on(struct acpi_power_resource *resource)ACPI_DEBUG_PRINT ((ACPI_DB_INFO ,"Power resource [%s] turned on\n" , resource->name)); +/* +* If there are other dependents on this power resource we need to +* resume them now so that their drivers can re-initialize the +* hardware properly after it went back to D0. +*/ +if (list_empty (&resource->dependents) || +list_is_singular (&resource->dependents)) +return 0 ; + +list_for_each_entry (dep, &resource->dependents, node) { +dev_dbg (dep->dev,"runtime resuming because [%s] turned on\n" , + resource->name); +pm_request_resume (dep->dev); + } +return 0 ; }
@@ -810,6 +948,7 @@ int acpi_add_power_resource(acpi_handle handle) ACPI_STA_DEFAULT);mutex_init (&resource->resource_lock);INIT_LIST_HEAD (&resource->list_node); +INIT_LIST_HEAD (&resource->dependents); resource->name = device->pnp.bus_id;strcpy (acpi_device_name (device),ACPI_POWER_DEVICE_NAME );strcpy (acpi_device_class (device),ACPI_POWER_CLASS );
include/acpi/acpi_bus.h +4 -0
@@ -513,6 +513,10 @@ int acpi_device_fix_up_power(struct acpi_device *device);int acpi_bus_update_power (acpi_handle handle,int *state_p);int acpi_device_update_power (struct acpi_device *device,int *state_p);bool acpi_bus_power_manageable (acpi_handle handle); +int acpi_device_power_add_dependent (struct acpi_device *adev, +struct device *dev); +void acpi_device_power_remove_dependent (struct acpi_device *adev, +struct device *dev);#ifdef CONFIG_PM bool acpi_bus_can_wakeup (acpi_handle handle);
Signature · 1 line
2.20.1
Rafael J. Wysocki
maintainerRe: [PATCH v2 2/3] ACPI / PM: Introduce concept of a _PR0 dependent device
Quoted 119 lines · Mika Westerberg
If there are shared power resources between otherwise unrelated devices turning them on causes the other devices sharing them to be powered up as well. In case of PCI devices go into D0uninitialized state meaning that if they were configured to trigger wake that configuration is lost at this point.
For this reason introduce a concept of "
_PR0dependent device" that can be added to any ACPI device that has power resources. The dependent device will be included in a list of dependent devices for all power resources returned by the ACPI device's_PR0(assuming it has one). Whenever a power resource having dependent devices is turned physically on (its_ONmethod is called) we runtime resume all of them to allow their driver or in case of PCI the PCI core to re-initialize the device and its wake configuration.This adds two functions that can be used to add and remove these dependent devices. Note the dependent device does not necessary need share power resources so this functionality can be used to add "software dependencies" as well if needed.
Signed-off-by Mika Westerberg <[email protected]>--- drivers/acpi/power.c | 139 ++++++++++++++++++++++++++++++++++++++++ include/acpi/acpi_bus.h | 4 ++ 2 files changed, 143 insertions(+)drivers/acpi/power.c +70
-0@@ -42,6 +42,11 @@ ACPI_MODULE_NAME("power");#define ACPI_POWER_RESOURCE_STATE_ON 0x01#define ACPI_POWER_RESOURCE_STATE_UNKNOWN 0xFF +struct acpi_power_dependent_device { +struct device *dev; +struct list_head node; +}; +struct acpi_power_resource {struct acpi_device device;struct list_head list_node;@@ -51,6 +56,7 @@ struct acpi_power_resource {unsigned int ref_count;bool wakeup_enabled;struct mutex resource_lock; +struct list_head dependents; };struct acpi_power_resource_entry {@@ -232,8 +238,125 @@ static int acpi_power_get_list_state(struct list_head *list, int *state)return 0 ; } +static int +acpi_power_resource_add_dependent (struct acpi_power_resource *resource, +struct device *dev) +{ +struct acpi_power_dependent_device *dep; +int ret =0 ; + +mutex_lock (&resource->resource_lock); +list_for_each_entry (dep, &resource->dependents, node) { +/* Only add it once */ +if (dep->dev == dev) + goto unlock; + } + + dep =kzalloc (sizeof (*dep),GFP_KERNEL ); +if (!dep) { + ret = -ENOMEM ; + goto unlock; + } + + dep->dev = dev; +list_add_tail (&dep->node, &resource->dependents); +dev_dbg (dev,"added power dependency to [%s]\n" , resource->name); + +unlock: +mutex_unlock (&resource->resource_lock); +return ret; +} + +static void +acpi_power_resource_remove_dependent (struct acpi_power_resource *resource, +struct device *dev) +{ +struct acpi_power_dependent_device *dep; + +mutex_lock (&resource->resource_lock); +list_for_each_entry (dep, &resource->dependents, node) { +if (dep->dev == dev) { +list_del (&dep->node); +kfree (dep); +dev_dbg (dev,"removed power dependency to [%s]\n" , + resource->name); +break ; + } + } +mutex_unlock (&resource->resource_lock); +} + +/** +* acpi_device_power_add_dependent - Add dependent device of this ACPI device +* @adev: ACPI device pointer +* @dev: Dependent device +* +* If @adev has non-empty _PR0 the @dev is added as dependent device to all +* power resources returned by it. This means that whenever these power +* resources are turned _ON the dependent devices get runtime resumed. This +* is needed for devices such as PCI to allow its driver to re-initialize +* it after it went to D0uninitialized. +* +* If @adev does not have _PR0 this does nothing. +* +* Returns %0 in case of success and negative errno otherwise. +*/ +int acpi_device_power_add_dependent(struct acpi_device *adev,
drivers/acpi/power.c static int acpi_power_get_list_state(struct list_head *list, int *state) · line 314
Quoted 10 lines
drivers/acpi/power.c +10
-0+ struct device *dev) +{ +struct acpi_power_resource_entry *entry; +struct list_head *resources; +int ret; + +if (!adev->power.flags.power_resources) +return 0 ; +if (!adev->power.states[ACPI_STATE_D0 ].flags.valid) +return 0 ;
The two checks above can be replaced with an adev->flags.power_manageable one AFAICS (the "valid" flag is always set for D0 and the list below will be empty if there are no power resources).
Same for acpi_device_power_remove_dependent(), of course.
Apart from this LGTM.
Re: [PATCH v2 2/3] ACPI / PM: Introduce concept of a _PR0 dependent device
Quoted 17 lines · Rafael J. Wysocki
drivers/acpi/power.c static int acpi_power_get_list_state(struct list_head *list, int *state) · line 314
Quoted 11 lines
drivers/acpi/power.c +11
-0+int acpi_device_power_add_dependent (struct acpi_device *adev, +struct device *dev) +{ +struct acpi_power_resource_entry *entry; +struct list_head *resources; +int ret; + +if (!adev->power.flags.power_resources) +return 0 ; +if (!adev->power.states[ACPI_STATE_D0 ].flags.valid) +return 0 ;The two checks above can be replaced with an
adev->flags.power_manageableone AFAICS (the "valid" flag is always set for D0 and the list below will be empty if there are no power resources).
Quoted 1 line
Same for
acpi_device_power_remove_dependent(), of course.
OK, I'll do that in next version.
Quoted 1 line
Apart from this LGTM.
Thanks!
PATCH v2 3/3 PCI / ACPI: Add _PR0 dependent devices
In reply to Mika WesterbergIf otherwise unrelated PCI devices share ACPI power resources turning them on causes the devices to enter D0uninitialized power state which may cause problems.
For example in Intel Ice Lake two root ports (RP0 and RP1), Thunderbolt controller (NHI) and xHCI controller all share power resources as can be ween in the topology below where power resources are marked with []:
Host bridge
|
+- RP0 ---\
+- RP1 ---|--+--> [TBT]
+- NHI --/ |
| |
| v
+- xHCI --> [D3C]In a situation where all devices sharing the power resources are in D3cold (the power resources are turned off) and for example the Thunderbolt controller is runtime resumed resulting that the power resources are turned on. This means that the other devices sharing them (RP0, RP1 and xHCI) are transitioned into D0uninitialized state. If they were configured to trigger wake (PME) on a certain event that configuration gets lost after reset so we would need to re-initialize them to get the wakeup working as expected again. To do so we would need to runtime resume all of them to make sure their registers get restored properly before we can runtime suspend them again.
Since we just added concept of "_PR0 dependent device" we can solve this by calling the relevant add/remove functions when the PCI device is bind to its ACPI representation. If it has power resources the PCI device will be added as dependent device to them and runtime resumed whenever they are physically turned on. This should make sure PCI core can reconfigure wakes after the device is transitioned into D0uninitialized.
--- drivers/pci/pci-acpi.c | 2 ++ 1 file changed, 2 insertions(+)
drivers/pci/pci-acpi.c +2 -0
@@ -902,6 +902,7 @@ static void pci_acpi_setup(struct device *dev)device_wakeup_enable (dev);acpi_pci_wakeup (pci_dev, false); +acpi_device_power_add_dependent (adev, dev); }static void pci_acpi_cleanup (struct device *dev)
@@ -914,6 +915,7 @@ static void pci_acpi_cleanup(struct device *dev)pci_acpi_remove_pm_notifier (adev);if (adev->wakeup.flags.valid) { +acpi_device_power_remove_dependent (adev, dev);if (pci_dev->bridge_d3)device_wakeup_disable (dev);
Signature · 1 line
2.20.1
Rafael J. Wysocki
maintainerQuoted 31 lines · Mika Westerberg
Hi all,
Based on a discussion regarding patch series I sent previously [1] to deal with sibling devices sharing ACPI power resources, I prepared a new reworked version according to the comments I got.
To summarize, in Intel Ice Lake the Thunderbolt controller, PCIe root ports and xHCI all share power resources. When they are all in D3hot power resources (returned by
_PR3) can be turned off powering off the whole block. However, there are two issues around this.Firstly the PCI core sets the device power state by asking what the real ACPI power state is. This results that all but last device sharing the power resources are in D3hot when the power resources are turned off. This causes issues if user runs for example 'lspci' because the device is really in D3cold so what user gets back is all ones (
0xffffffff).Secondly if any of the device is runtime resumed the power resources are turned on bringing all other devices sharing the resources to D0uninitialized losing their wakeup configuration.
This series aims to fix the two issues by:
1. Using the ACPI cached power state when PCI devices are transitioned into low power states instead of reading back the "real" power state.
2. Introducing concept of "
_PR0dependent devices" that get runtime resumed whenever their power resource (which they might share with other sibling devices) gets turned on.
Quoted 8 lines
The series is based on the idea of Rafael J. Wysocki <[email protected]>.
[1] https://www.spinics.net/lists/linux-pci/msg83583.html
Mika Westerberg (3):
PCI / ACPI: Use cached ACPI device state to get PCI device power state ACPI / PM: Introduce concept of a_PR0dependent device PCI / ACPI: Add_PR0dependent devices
LGMT overall, patch [2/3] can be simplified slightly IMO (already sent comments for that one).
As far as I'm concerned, the other patches need not be updated.