Message ID | 20210617123653.58640-1-mika.westerberg@linux.intel.com (mailing list archive) |
---|---|
State | New, archived |
Delegated to: | Bjorn Helgaas |
Headers | show |
Series | [v4] PCI/PM: Target PM state is D3hot if device can only generate PME from D3cold | expand |
On Thu, Jun 17, 2021 at 2:36 PM Mika Westerberg <mika.westerberg@linux.intel.com> wrote: > > Some PCIe devices only support PME (Power Management Event) from D3cold. > One example is ASMedia xHCI controller: > > 11:00.0 USB controller: ASMedia Technology Inc. ASM1042A USB 3.0 Host Controller (prog-if 30 [XHCI]) > ... > Capabilities: [78] Power Management version 3 > Flags: PMEClk- DSI- D1- D2- AuxCurrent=55mA PME(D0-,D1-,D2-,D3hot-,D3cold+) > Status: D0 NoSoftRst+ PME-Enable- DSel=0 DScale=0 PME- > > With such devices, if it has wake enabled, the kernel selects lowest > possible power state to be D0 in pci_target_state(). This is problematic > because it prevents the root port it is connected to enter low power > state too which makes the system consume more energy than necessary. > > The problem in pci_target_state() is that it only accounts the "current" > device state, so when the bridge above it (a root port for instance) is > transitioned into D3hot the device transitions into D3cold. This is > because when the root port is first transitioned into D3hot then the > ACPI power resource is turned off which puts the PCIe link to L2/L3 (and > the root port and the device are in D3cold). If the root port is kept in > D3hot it still means that the device below it is still effectively in > D3cold as no configuration messages pass through. Furthermore the > implementation note of PCIe 5.0 sec 5.3.1.4 says that the device should > expect to be transitioned into D3cold soon after its link transitions > into L2/L3 Ready state. > > Taking the above into consideration, instead of forcing the device stay > in D0 we modify pci_target_state() to return D3hot in this special case > and make __pci_enable_wake() to enable PME too in this case. > > Reported-by: Utkarsh H Patel <utkarsh.h.patel@intel.com> > Reported-by: Koba Ko <koba.ko@canonical.com> > Suggested-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com> All of my comments have been addressed, so Reviewed-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > --- > The previous version of the patch is here: > > https://lore.kernel.org/linux-pm/20210616150516.28242-1-mika.westerberg@linux.intel.com/ > > Changes from the previous version: > > * Dropped redundant test in pci_target_state(). > > drivers/pci/pci.c | 16 +++++++++++++++- > 1 file changed, 15 insertions(+), 1 deletion(-) > > diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c > index b717680377a9..043c5c304308 100644 > --- a/drivers/pci/pci.c > +++ b/drivers/pci/pci.c > @@ -2485,7 +2485,13 @@ static int __pci_enable_wake(struct pci_dev *dev, pci_power_t state, bool enable > if (enable) { > int error; > > - if (pci_pme_capable(dev, state)) > + /* > + * Enable PME if device is capable from given state. > + * Special case is device that can only generate PME > + * from D3cold then we enable PME too. > + */ > + if (pci_pme_capable(dev, state) || > + (state == PCI_D3hot && pci_pme_capable(dev, PCI_D3cold))) > pci_pme_active(dev, true); > else > ret = 1; > @@ -2595,6 +2601,14 @@ static pci_power_t pci_target_state(struct pci_dev *dev, bool wakeup) > * PME#. > */ > if (dev->pme_support) { > + /* > + * Special case if device supports only PME from > + * D3cold but not from D3hot we still return D3hot. > + */ > + if (target_state == PCI_D3hot && > + (dev->pme_support & (1 << PCI_D3cold))) > + return target_state; > + > while (target_state > && !(dev->pme_support & (1 << target_state))) > target_state--; > -- > 2.30.2 >
On Thu, Jun 17, 2021 at 03:36:53PM +0300, Mika Westerberg wrote: > Some PCIe devices only support PME (Power Management Event) from D3cold. > One example is ASMedia xHCI controller: > > 11:00.0 USB controller: ASMedia Technology Inc. ASM1042A USB 3.0 Host Controller (prog-if 30 [XHCI]) > ... > Capabilities: [78] Power Management version 3 > Flags: PMEClk- DSI- D1- D2- AuxCurrent=55mA PME(D0-,D1-,D2-,D3hot-,D3cold+) > Status: D0 NoSoftRst+ PME-Enable- DSel=0 DScale=0 PME- > > With such devices, if it has wake enabled, the kernel selects lowest > possible power state to be D0 in pci_target_state(). This is problematic > because it prevents the root port it is connected to enter low power > state too which makes the system consume more energy than necessary. IIUC this is because the loop that checks which states support PME starts with D3hot and doesn't even look at D3cold. > The problem in pci_target_state() is that it only accounts the "current" > device state, so when the bridge above it (a root port for instance) is > transitioned into D3hot the device transitions into D3cold. This is > because when the root port is first transitioned into D3hot then the > ACPI power resource is turned off which puts the PCIe link to L2/L3 (and > the root port and the device are in D3cold). If the root port is kept in > D3hot it still means that the device below it is still effectively in > D3cold as no configuration messages pass through. Furthermore the > implementation note of PCIe 5.0 sec 5.3.1.4 says that the device should > expect to be transitioned into D3cold soon after its link transitions > into L2/L3 Ready state. > > Taking the above into consideration, instead of forcing the device stay > in D0 we modify pci_target_state() to return D3hot in this special case > and make __pci_enable_wake() to enable PME too in this case. > > Reported-by: Utkarsh H Patel <utkarsh.h.patel@intel.com> > Reported-by: Koba Ko <koba.ko@canonical.com> > Suggested-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com> > --- > The previous version of the patch is here: > > https://lore.kernel.org/linux-pm/20210616150516.28242-1-mika.westerberg@linux.intel.com/ > > Changes from the previous version: > > * Dropped redundant test in pci_target_state(). > > drivers/pci/pci.c | 16 +++++++++++++++- > 1 file changed, 15 insertions(+), 1 deletion(-) > > diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c > index b717680377a9..043c5c304308 100644 > --- a/drivers/pci/pci.c > +++ b/drivers/pci/pci.c > @@ -2485,7 +2485,13 @@ static int __pci_enable_wake(struct pci_dev *dev, pci_power_t state, bool enable > if (enable) { > int error; > > - if (pci_pme_capable(dev, state)) > + /* > + * Enable PME if device is capable from given state. > + * Special case is device that can only generate PME > + * from D3cold then we enable PME too. > + */ > + if (pci_pme_capable(dev, state) || > + (state == PCI_D3hot && pci_pme_capable(dev, PCI_D3cold))) > pci_pme_active(dev, true); > else > ret = 1; > @@ -2595,6 +2601,14 @@ static pci_power_t pci_target_state(struct pci_dev *dev, bool wakeup) > * PME#. > */ > if (dev->pme_support) { > + /* > + * Special case if device supports only PME from > + * D3cold but not from D3hot we still return D3hot. > + */ > + if (target_state == PCI_D3hot && > + (dev->pme_support & (1 << PCI_D3cold))) > + return target_state; I've spent quite a bit of time trying to understand this, and I'm kind of dragging my feet on it because I haven't been able to really connect this with the specs. It also seems unfortunate to have to add this special case in two places. It seems like we're basically lying and *saying* we're going to put the device in D3hot, but due to some magic invisible assumption, we *actually* put it in D3cold. > while (target_state > && !(dev->pme_support & (1 << target_state))) > target_state--; Nit 1: "if (target_state == PCI_D3hot && ...) return target_state;" means "if (...) return PCI_D3hot;". When we're returning a constant value that we already know, I think it's clearer to use the constant. Nit 2: it looks like both these tests should use pci_pme_capable(), which would match the other special case in __pci_enable_wake().
On Wed, Jul 7, 2021 at 11:57 PM Bjorn Helgaas <helgaas@kernel.org> wrote: > > On Thu, Jun 17, 2021 at 03:36:53PM +0300, Mika Westerberg wrote: > > Some PCIe devices only support PME (Power Management Event) from D3cold. > > One example is ASMedia xHCI controller: > > > > 11:00.0 USB controller: ASMedia Technology Inc. ASM1042A USB 3.0 Host Controller (prog-if 30 [XHCI]) > > ... > > Capabilities: [78] Power Management version 3 > > Flags: PMEClk- DSI- D1- D2- AuxCurrent=55mA PME(D0-,D1-,D2-,D3hot-,D3cold+) > > Status: D0 NoSoftRst+ PME-Enable- DSel=0 DScale=0 PME- > > > > With such devices, if it has wake enabled, the kernel selects lowest > > possible power state to be D0 in pci_target_state(). This is problematic > > because it prevents the root port it is connected to enter low power > > state too which makes the system consume more energy than necessary. > > IIUC this is because the loop that checks which states support PME > starts with D3hot and doesn't even look at D3cold. That's because the device itself cannot be programmed into D3cold, so the target state cannot be D3cold for it. > > The problem in pci_target_state() is that it only accounts the "current" > > device state, so when the bridge above it (a root port for instance) is > > transitioned into D3hot the device transitions into D3cold. This is > > because when the root port is first transitioned into D3hot then the > > ACPI power resource is turned off which puts the PCIe link to L2/L3 (and > > the root port and the device are in D3cold). If the root port is kept in > > D3hot it still means that the device below it is still effectively in > > D3cold as no configuration messages pass through. Furthermore the > > implementation note of PCIe 5.0 sec 5.3.1.4 says that the device should > > expect to be transitioned into D3cold soon after its link transitions > > into L2/L3 Ready state. > > > > Taking the above into consideration, instead of forcing the device stay > > in D0 we modify pci_target_state() to return D3hot in this special case > > and make __pci_enable_wake() to enable PME too in this case. > > > > Reported-by: Utkarsh H Patel <utkarsh.h.patel@intel.com> > > Reported-by: Koba Ko <koba.ko@canonical.com> > > Suggested-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > > Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com> > > --- > > The previous version of the patch is here: > > > > https://lore.kernel.org/linux-pm/20210616150516.28242-1-mika.westerberg@linux.intel.com/ > > > > Changes from the previous version: > > > > * Dropped redundant test in pci_target_state(). > > > > drivers/pci/pci.c | 16 +++++++++++++++- > > 1 file changed, 15 insertions(+), 1 deletion(-) > > > > diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c > > index b717680377a9..043c5c304308 100644 > > --- a/drivers/pci/pci.c > > +++ b/drivers/pci/pci.c > > @@ -2485,7 +2485,13 @@ static int __pci_enable_wake(struct pci_dev *dev, pci_power_t state, bool enable > > if (enable) { > > int error; > > > > - if (pci_pme_capable(dev, state)) > > + /* > > + * Enable PME if device is capable from given state. > > + * Special case is device that can only generate PME > > + * from D3cold then we enable PME too. > > + */ > > + if (pci_pme_capable(dev, state) || > > + (state == PCI_D3hot && pci_pme_capable(dev, PCI_D3cold))) > > pci_pme_active(dev, true); > > else > > ret = 1; > > @@ -2595,6 +2601,14 @@ static pci_power_t pci_target_state(struct pci_dev *dev, bool wakeup) > > * PME#. > > */ > > if (dev->pme_support) { > > + /* > > + * Special case if device supports only PME from > > + * D3cold but not from D3hot we still return D3hot. > > + */ > > + if (target_state == PCI_D3hot && > > + (dev->pme_support & (1 << PCI_D3cold))) > > + return target_state; > > I've spent quite a bit of time trying to understand this, and I'm kind > of dragging my feet on it because I haven't been able to really > connect this with the specs. The specs aren't very clear in this area, though. The overall picture is that the device in question is connected to a port (a root port in this particular case) that can be programmed into D3cold via ACPI, but the endpoint itself can only be programmed into D3hot. However, if the port goes into D3cold, the endpoint also goes into D3cold (actually, my understanding of the specs is that even if the port goes into D3hot, the endpoint should still be assumed to go into D3cold). The power state of the endpoint is changed first and at the time this happens it is not known which power state the port is going to be programmed into. Now, the device is wake-capable (in general) and so we want it to be able to signal wakeup from the final power state. Because it only reports PME support in D0 and in D3cold, the kernel today leaves it in D0 which causes the port to stay in D0 too. Still, putting the device into D3hot allows the port to go into D3cold which in turn causes the device to go into D3cold and it can signal wakeup from that state. So there are two ways to get into a configuration from which the endpoint device can signal wakeup, either by leaving it and the port holding it both in D0, or by putting it into D3hot, so that the port can go into D3cold in which case the endpoint will end up in D3cold. The Mika's patch is aiming at enabling the second option. > It also seems unfortunate to have to add this special case in two places. That's because __pci_enable_wake() tries to be extra careful and only call pci_pme_active() if PME is known to be supported in the target power state, but that is not strictly necessary. It could just call pci_pme_active() unconditionally and return the platform_pci_set_wakeup() return value. I think I'll send a patch making this change. > It seems like we're basically lying and *saying* we're going to put > the device in D3hot, but due to some magic invisible assumption, we > *actually* put it in D3cold. Not really. We put the device indo D3hot, because it will enable its parent port to go into D3cold and that will cause the device itself to end up in D3cold. It looks now like the parent's capability to go into D3cold should be checked here. > > while (target_state > > && !(dev->pme_support & (1 << target_state))) > > target_state--; > > Nit 1: "if (target_state == PCI_D3hot && ...) return target_state;" > means "if (...) return PCI_D3hot;". When we're returning a constant > value that we already know, I think it's clearer to use the constant. Good point. > Nit 2: it looks like both these tests should use pci_pme_capable(), > which would match the other special case in __pci_enable_wake(). That would be cleaner.
On Thu, Jul 8, 2021 at 2:18 PM Rafael J. Wysocki <rafael@kernel.org> wrote: > > On Wed, Jul 7, 2021 at 11:57 PM Bjorn Helgaas <helgaas@kernel.org> wrote: > > > > On Thu, Jun 17, 2021 at 03:36:53PM +0300, Mika Westerberg wrote: > > > Some PCIe devices only support PME (Power Management Event) from D3cold. > > > One example is ASMedia xHCI controller: > > > > > > 11:00.0 USB controller: ASMedia Technology Inc. ASM1042A USB 3.0 Host Controller (prog-if 30 [XHCI]) > > > ... > > > Capabilities: [78] Power Management version 3 > > > Flags: PMEClk- DSI- D1- D2- AuxCurrent=55mA PME(D0-,D1-,D2-,D3hot-,D3cold+) > > > Status: D0 NoSoftRst+ PME-Enable- DSel=0 DScale=0 PME- > > > > > > With such devices, if it has wake enabled, the kernel selects lowest > > > possible power state to be D0 in pci_target_state(). This is problematic > > > because it prevents the root port it is connected to enter low power > > > state too which makes the system consume more energy than necessary. > > > > IIUC this is because the loop that checks which states support PME > > starts with D3hot and doesn't even look at D3cold. > > That's because the device itself cannot be programmed into D3cold, so > the target state cannot be D3cold for it. > > > > The problem in pci_target_state() is that it only accounts the "current" > > > device state, so when the bridge above it (a root port for instance) is > > > transitioned into D3hot the device transitions into D3cold. This is > > > because when the root port is first transitioned into D3hot then the > > > ACPI power resource is turned off which puts the PCIe link to L2/L3 (and > > > the root port and the device are in D3cold). If the root port is kept in > > > D3hot it still means that the device below it is still effectively in > > > D3cold as no configuration messages pass through. Furthermore the > > > implementation note of PCIe 5.0 sec 5.3.1.4 says that the device should > > > expect to be transitioned into D3cold soon after its link transitions > > > into L2/L3 Ready state. > > > > > > Taking the above into consideration, instead of forcing the device stay > > > in D0 we modify pci_target_state() to return D3hot in this special case > > > and make __pci_enable_wake() to enable PME too in this case. > > > > > > Reported-by: Utkarsh H Patel <utkarsh.h.patel@intel.com> > > > Reported-by: Koba Ko <koba.ko@canonical.com> > > > Suggested-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > > > Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com> > > > --- > > > The previous version of the patch is here: > > > > > > https://lore.kernel.org/linux-pm/20210616150516.28242-1-mika.westerberg@linux.intel.com/ > > > > > > Changes from the previous version: > > > > > > * Dropped redundant test in pci_target_state(). > > > > > > drivers/pci/pci.c | 16 +++++++++++++++- > > > 1 file changed, 15 insertions(+), 1 deletion(-) > > > > > > diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c > > > index b717680377a9..043c5c304308 100644 > > > --- a/drivers/pci/pci.c > > > +++ b/drivers/pci/pci.c > > > @@ -2485,7 +2485,13 @@ static int __pci_enable_wake(struct pci_dev *dev, pci_power_t state, bool enable > > > if (enable) { > > > int error; > > > > > > - if (pci_pme_capable(dev, state)) > > > + /* > > > + * Enable PME if device is capable from given state. > > > + * Special case is device that can only generate PME > > > + * from D3cold then we enable PME too. > > > + */ > > > + if (pci_pme_capable(dev, state) || > > > + (state == PCI_D3hot && pci_pme_capable(dev, PCI_D3cold))) > > > pci_pme_active(dev, true); > > > else > > > ret = 1; > > > @@ -2595,6 +2601,14 @@ static pci_power_t pci_target_state(struct pci_dev *dev, bool wakeup) > > > * PME#. > > > */ > > > if (dev->pme_support) { > > > + /* > > > + * Special case if device supports only PME from > > > + * D3cold but not from D3hot we still return D3hot. > > > + */ > > > + if (target_state == PCI_D3hot && > > > + (dev->pme_support & (1 << PCI_D3cold))) > > > + return target_state; > > > > I've spent quite a bit of time trying to understand this, and I'm kind > > of dragging my feet on it because I haven't been able to really > > connect this with the specs. > > The specs aren't very clear in this area, though. > > The overall picture is that the device in question is connected to a > port (a root port in this particular case) that can be programmed into > D3cold via ACPI, but the endpoint itself can only be programmed into > D3hot. However, if the port goes into D3cold, the endpoint also goes > into D3cold (actually, my understanding of the specs is that even if > the port goes into D3hot, the endpoint should still be assumed to go > into D3cold). > > The power state of the endpoint is changed first and at the time this > happens it is not known which power state the port is going to be > programmed into. > > Now, the device is wake-capable (in general) and so we want it to be > able to signal wakeup from the final power state. Because it only > reports PME support in D0 and in D3cold, the kernel today leaves it in > D0 which causes the port to stay in D0 too. Still, putting the device > into D3hot allows the port to go into D3cold which in turn causes the > device to go into D3cold and it can signal wakeup from that state. > > So there are two ways to get into a configuration from which the > endpoint device can signal wakeup, either by leaving it and the port > holding it both in D0, or by putting it into D3hot, so that the port > can go into D3cold in which case the endpoint will end up in D3cold. > > The Mika's patch is aiming at enabling the second option. > > > It also seems unfortunate to have to add this special case in two places. > > That's because __pci_enable_wake() tries to be extra careful and only > call pci_pme_active() if PME is known to be supported in the target > power state, but that is not strictly necessary. It could just call > pci_pme_active() unconditionally and return the > platform_pci_set_wakeup() return value. > > I think I'll send a patch making this change. Actually, it needs to fail if PME cannot be signaled from the target state and the device is not power-manageable by the platform. A better idea may be to make pci_pme_capable() also check if the parent bridge can go into D3cold and return "true" if so and "state" is D3hot while PME signaling from D3cold is supported.
On Thursday, July 8, 2021 2:39:49 PM CEST Rafael J. Wysocki wrote: > On Thu, Jul 8, 2021 at 2:18 PM Rafael J. Wysocki <rafael@kernel.org> wrote: > > > > On Wed, Jul 7, 2021 at 11:57 PM Bjorn Helgaas <helgaas@kernel.org> wrote: > > > > > > On Thu, Jun 17, 2021 at 03:36:53PM +0300, Mika Westerberg wrote: > > > > Some PCIe devices only support PME (Power Management Event) from D3cold. > > > > One example is ASMedia xHCI controller: > > > > > > > > 11:00.0 USB controller: ASMedia Technology Inc. ASM1042A USB 3.0 Host Controller (prog-if 30 [XHCI]) > > > > ... > > > > Capabilities: [78] Power Management version 3 > > > > Flags: PMEClk- DSI- D1- D2- AuxCurrent=55mA PME(D0-,D1-,D2-,D3hot-,D3cold+) > > > > Status: D0 NoSoftRst+ PME-Enable- DSel=0 DScale=0 PME- > > > > > > > > With such devices, if it has wake enabled, the kernel selects lowest > > > > possible power state to be D0 in pci_target_state(). This is problematic > > > > because it prevents the root port it is connected to enter low power > > > > state too which makes the system consume more energy than necessary. > > > > > > IIUC this is because the loop that checks which states support PME > > > starts with D3hot and doesn't even look at D3cold. > > > > That's because the device itself cannot be programmed into D3cold, so > > the target state cannot be D3cold for it. > > > > > > The problem in pci_target_state() is that it only accounts the "current" > > > > device state, so when the bridge above it (a root port for instance) is > > > > transitioned into D3hot the device transitions into D3cold. This is > > > > because when the root port is first transitioned into D3hot then the > > > > ACPI power resource is turned off which puts the PCIe link to L2/L3 (and > > > > the root port and the device are in D3cold). If the root port is kept in > > > > D3hot it still means that the device below it is still effectively in > > > > D3cold as no configuration messages pass through. Furthermore the > > > > implementation note of PCIe 5.0 sec 5.3.1.4 says that the device should > > > > expect to be transitioned into D3cold soon after its link transitions > > > > into L2/L3 Ready state. > > > > > > > > Taking the above into consideration, instead of forcing the device stay > > > > in D0 we modify pci_target_state() to return D3hot in this special case > > > > and make __pci_enable_wake() to enable PME too in this case. > > > > > > > > Reported-by: Utkarsh H Patel <utkarsh.h.patel@intel.com> > > > > Reported-by: Koba Ko <koba.ko@canonical.com> > > > > Suggested-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > > > > Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com> > > > > --- > > > > The previous version of the patch is here: > > > > > > > > https://lore.kernel.org/linux-pm/20210616150516.28242-1-mika.westerberg@linux.intel.com/ > > > > > > > > Changes from the previous version: > > > > > > > > * Dropped redundant test in pci_target_state(). > > > > > > > > drivers/pci/pci.c | 16 +++++++++++++++- > > > > 1 file changed, 15 insertions(+), 1 deletion(-) > > > > > > > > diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c > > > > index b717680377a9..043c5c304308 100644 > > > > --- a/drivers/pci/pci.c > > > > +++ b/drivers/pci/pci.c > > > > @@ -2485,7 +2485,13 @@ static int __pci_enable_wake(struct pci_dev *dev, pci_power_t state, bool enable > > > > if (enable) { > > > > int error; > > > > > > > > - if (pci_pme_capable(dev, state)) > > > > + /* > > > > + * Enable PME if device is capable from given state. > > > > + * Special case is device that can only generate PME > > > > + * from D3cold then we enable PME too. > > > > + */ > > > > + if (pci_pme_capable(dev, state) || > > > > + (state == PCI_D3hot && pci_pme_capable(dev, PCI_D3cold))) > > > > pci_pme_active(dev, true); > > > > else > > > > ret = 1; > > > > @@ -2595,6 +2601,14 @@ static pci_power_t pci_target_state(struct pci_dev *dev, bool wakeup) > > > > * PME#. > > > > */ > > > > if (dev->pme_support) { > > > > + /* > > > > + * Special case if device supports only PME from > > > > + * D3cold but not from D3hot we still return D3hot. > > > > + */ > > > > + if (target_state == PCI_D3hot && > > > > + (dev->pme_support & (1 << PCI_D3cold))) > > > > + return target_state; > > > > > > I've spent quite a bit of time trying to understand this, and I'm kind > > > of dragging my feet on it because I haven't been able to really > > > connect this with the specs. > > > > The specs aren't very clear in this area, though. > > > > The overall picture is that the device in question is connected to a > > port (a root port in this particular case) that can be programmed into > > D3cold via ACPI, but the endpoint itself can only be programmed into > > D3hot. However, if the port goes into D3cold, the endpoint also goes > > into D3cold (actually, my understanding of the specs is that even if > > the port goes into D3hot, the endpoint should still be assumed to go > > into D3cold). > > > > The power state of the endpoint is changed first and at the time this > > happens it is not known which power state the port is going to be > > programmed into. > > > > Now, the device is wake-capable (in general) and so we want it to be > > able to signal wakeup from the final power state. Because it only > > reports PME support in D0 and in D3cold, the kernel today leaves it in > > D0 which causes the port to stay in D0 too. Still, putting the device > > into D3hot allows the port to go into D3cold which in turn causes the > > device to go into D3cold and it can signal wakeup from that state. > > > > So there are two ways to get into a configuration from which the > > endpoint device can signal wakeup, either by leaving it and the port > > holding it both in D0, or by putting it into D3hot, so that the port > > can go into D3cold in which case the endpoint will end up in D3cold. > > > > The Mika's patch is aiming at enabling the second option. > > > > > It also seems unfortunate to have to add this special case in two places. > > > > That's because __pci_enable_wake() tries to be extra careful and only > > call pci_pme_active() if PME is known to be supported in the target > > power state, but that is not strictly necessary. It could just call > > pci_pme_active() unconditionally and return the > > platform_pci_set_wakeup() return value. > > > > I think I'll send a patch making this change. > > Actually, it needs to fail if PME cannot be signaled from the target > state and the device is not power-manageable by the platform. > > A better idea may be to make pci_pme_capable() also check if the > parent bridge can go into D3cold and return "true" if so and "state" > is D3hot while PME signaling from D3cold is supported. So below is my version of the $subject patch (untested). Please let me know what you think. --- drivers/pci/pci.c | 31 +++++++++++++++++++------------ 1 file changed, 19 insertions(+), 12 deletions(-) Index: linux-pm/drivers/pci/pci.c =================================================================== --- linux-pm.orig/drivers/pci/pci.c +++ linux-pm/drivers/pci/pci.c @@ -2305,7 +2305,19 @@ bool pci_pme_capable(struct pci_dev *dev if (!dev->pm_cap) return false; - return !!(dev->pme_support & (1 << state)); + if (dev->pme_support & (1 << state)) + return true; + + if (state < PCI_D3hot) + return false; + + /* + * If the device goes into D3hot, the parent bridge is allowed to go + * into D3 and the device will end up in D3cold, so if it supports + * signaling PME from D3cold, it still should be good then. + */ + return pci_bridge_d3_possible(dev->bus->self) && + (dev->pme_support & (1 << PCI_D3cold)); } EXPORT_SYMBOL(pci_pme_capable); @@ -2599,17 +2611,12 @@ static pci_power_t pci_target_state(stru if (dev->current_state == PCI_D3cold) target_state = PCI_D3cold; - if (wakeup) { - /* - * Find the deepest state from which the device can generate - * PME#. - */ - if (dev->pme_support) { - while (target_state - && !(dev->pme_support & (1 << target_state))) - target_state--; - } - } + if (!wakeup || pci_pme_capable(dev, target_state) || !dev->pme_support) + return target_state; + + /* Find the deepest state from which the device can generate PME#. */ + while (target_state && !(dev->pme_support & (1 << target_state))) + target_state--; return target_state; }
On Thu, Jul 8, 2021 at 3:20 PM Rafael J. Wysocki <rjw@rjwysocki.net> wrote: > > On Thursday, July 8, 2021 2:39:49 PM CEST Rafael J. Wysocki wrote: > > On Thu, Jul 8, 2021 at 2:18 PM Rafael J. Wysocki <rafael@kernel.org> wrote: > > > > > > On Wed, Jul 7, 2021 at 11:57 PM Bjorn Helgaas <helgaas@kernel.org> wrote: > > > > > > > > On Thu, Jun 17, 2021 at 03:36:53PM +0300, Mika Westerberg wrote: > > > > > Some PCIe devices only support PME (Power Management Event) from D3cold. > > > > > One example is ASMedia xHCI controller: > > > > > > > > > > 11:00.0 USB controller: ASMedia Technology Inc. ASM1042A USB 3.0 Host Controller (prog-if 30 [XHCI]) > > > > > ... > > > > > Capabilities: [78] Power Management version 3 > > > > > Flags: PMEClk- DSI- D1- D2- AuxCurrent=55mA PME(D0-,D1-,D2-,D3hot-,D3cold+) > > > > > Status: D0 NoSoftRst+ PME-Enable- DSel=0 DScale=0 PME- > > > > > > > > > > With such devices, if it has wake enabled, the kernel selects lowest > > > > > possible power state to be D0 in pci_target_state(). This is problematic > > > > > because it prevents the root port it is connected to enter low power > > > > > state too which makes the system consume more energy than necessary. > > > > > > > > IIUC this is because the loop that checks which states support PME > > > > starts with D3hot and doesn't even look at D3cold. > > > > > > That's because the device itself cannot be programmed into D3cold, so > > > the target state cannot be D3cold for it. > > > > > > > > The problem in pci_target_state() is that it only accounts the "current" > > > > > device state, so when the bridge above it (a root port for instance) is > > > > > transitioned into D3hot the device transitions into D3cold. This is > > > > > because when the root port is first transitioned into D3hot then the > > > > > ACPI power resource is turned off which puts the PCIe link to L2/L3 (and > > > > > the root port and the device are in D3cold). If the root port is kept in > > > > > D3hot it still means that the device below it is still effectively in > > > > > D3cold as no configuration messages pass through. Furthermore the > > > > > implementation note of PCIe 5.0 sec 5.3.1.4 says that the device should > > > > > expect to be transitioned into D3cold soon after its link transitions > > > > > into L2/L3 Ready state. > > > > > > > > > > Taking the above into consideration, instead of forcing the device stay > > > > > in D0 we modify pci_target_state() to return D3hot in this special case > > > > > and make __pci_enable_wake() to enable PME too in this case. > > > > > > > > > > Reported-by: Utkarsh H Patel <utkarsh.h.patel@intel.com> > > > > > Reported-by: Koba Ko <koba.ko@canonical.com> > > > > > Suggested-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > > > > > Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com> > > > > > --- > > > > > The previous version of the patch is here: > > > > > > > > > > https://lore.kernel.org/linux-pm/20210616150516.28242-1-mika.westerberg@linux.intel.com/ > > > > > > > > > > Changes from the previous version: > > > > > > > > > > * Dropped redundant test in pci_target_state(). > > > > > > > > > > drivers/pci/pci.c | 16 +++++++++++++++- > > > > > 1 file changed, 15 insertions(+), 1 deletion(-) > > > > > > > > > > diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c > > > > > index b717680377a9..043c5c304308 100644 > > > > > --- a/drivers/pci/pci.c > > > > > +++ b/drivers/pci/pci.c > > > > > @@ -2485,7 +2485,13 @@ static int __pci_enable_wake(struct pci_dev *dev, pci_power_t state, bool enable > > > > > if (enable) { > > > > > int error; > > > > > > > > > > - if (pci_pme_capable(dev, state)) > > > > > + /* > > > > > + * Enable PME if device is capable from given state. > > > > > + * Special case is device that can only generate PME > > > > > + * from D3cold then we enable PME too. > > > > > + */ > > > > > + if (pci_pme_capable(dev, state) || > > > > > + (state == PCI_D3hot && pci_pme_capable(dev, PCI_D3cold))) > > > > > pci_pme_active(dev, true); > > > > > else > > > > > ret = 1; > > > > > @@ -2595,6 +2601,14 @@ static pci_power_t pci_target_state(struct pci_dev *dev, bool wakeup) > > > > > * PME#. > > > > > */ > > > > > if (dev->pme_support) { > > > > > + /* > > > > > + * Special case if device supports only PME from > > > > > + * D3cold but not from D3hot we still return D3hot. > > > > > + */ > > > > > + if (target_state == PCI_D3hot && > > > > > + (dev->pme_support & (1 << PCI_D3cold))) > > > > > + return target_state; > > > > > > > > I've spent quite a bit of time trying to understand this, and I'm kind > > > > of dragging my feet on it because I haven't been able to really > > > > connect this with the specs. > > > > > > The specs aren't very clear in this area, though. > > > > > > The overall picture is that the device in question is connected to a > > > port (a root port in this particular case) that can be programmed into > > > D3cold via ACPI, but the endpoint itself can only be programmed into > > > D3hot. However, if the port goes into D3cold, the endpoint also goes > > > into D3cold (actually, my understanding of the specs is that even if > > > the port goes into D3hot, the endpoint should still be assumed to go > > > into D3cold). > > > > > > The power state of the endpoint is changed first and at the time this > > > happens it is not known which power state the port is going to be > > > programmed into. > > > > > > Now, the device is wake-capable (in general) and so we want it to be > > > able to signal wakeup from the final power state. Because it only > > > reports PME support in D0 and in D3cold, the kernel today leaves it in > > > D0 which causes the port to stay in D0 too. Still, putting the device > > > into D3hot allows the port to go into D3cold which in turn causes the > > > device to go into D3cold and it can signal wakeup from that state. > > > > > > So there are two ways to get into a configuration from which the > > > endpoint device can signal wakeup, either by leaving it and the port > > > holding it both in D0, or by putting it into D3hot, so that the port > > > can go into D3cold in which case the endpoint will end up in D3cold. > > > > > > The Mika's patch is aiming at enabling the second option. > > > > > > > It also seems unfortunate to have to add this special case in two places. > > > > > > That's because __pci_enable_wake() tries to be extra careful and only > > > call pci_pme_active() if PME is known to be supported in the target > > > power state, but that is not strictly necessary. It could just call > > > pci_pme_active() unconditionally and return the > > > platform_pci_set_wakeup() return value. > > > > > > I think I'll send a patch making this change. > > > > Actually, it needs to fail if PME cannot be signaled from the target > > state and the device is not power-manageable by the platform. > > > > A better idea may be to make pci_pme_capable() also check if the > > parent bridge can go into D3cold and return "true" if so and "state" > > is D3hot while PME signaling from D3cold is supported. > > So below is my version of the $subject patch (untested). > > Please let me know what you think. I gave some more consideration to this and I was not able to convince myself that putting the parent port into D3hot was sufficient for the endpoint device connected to it to go into D3cold. However, the PCI PM spec v1.2 clearly mandates that putting a bridge into D3cold will cause power to be removed from the entire bus segment below it, which should apply to PCIe devices by extension. So I'm going to submit a new version of the patch below in which pci_pme_capable() will check whether or not the parent will go into D3cold in addition to all of the other checks. Stay tuned! > > --- > drivers/pci/pci.c | 31 +++++++++++++++++++------------ > 1 file changed, 19 insertions(+), 12 deletions(-) > > Index: linux-pm/drivers/pci/pci.c > =================================================================== > --- linux-pm.orig/drivers/pci/pci.c > +++ linux-pm/drivers/pci/pci.c > @@ -2305,7 +2305,19 @@ bool pci_pme_capable(struct pci_dev *dev > if (!dev->pm_cap) > return false; > > - return !!(dev->pme_support & (1 << state)); > + if (dev->pme_support & (1 << state)) > + return true; > + > + if (state < PCI_D3hot) > + return false; > + > + /* > + * If the device goes into D3hot, the parent bridge is allowed to go > + * into D3 and the device will end up in D3cold, so if it supports > + * signaling PME from D3cold, it still should be good then. > + */ > + return pci_bridge_d3_possible(dev->bus->self) && > + (dev->pme_support & (1 << PCI_D3cold)); > } > EXPORT_SYMBOL(pci_pme_capable); > > @@ -2599,17 +2611,12 @@ static pci_power_t pci_target_state(stru > if (dev->current_state == PCI_D3cold) > target_state = PCI_D3cold; > > - if (wakeup) { > - /* > - * Find the deepest state from which the device can generate > - * PME#. > - */ > - if (dev->pme_support) { > - while (target_state > - && !(dev->pme_support & (1 << target_state))) > - target_state--; > - } > - } > + if (!wakeup || pci_pme_capable(dev, target_state) || !dev->pme_support) > + return target_state; > + > + /* Find the deepest state from which the device can generate PME#. */ > + while (target_state && !(dev->pme_support & (1 << target_state))) > + target_state--; > > return target_state; > } > > > >
On Mon, Jul 12, 2021 at 05:22:58PM +0200, Rafael J. Wysocki wrote: > On Thu, Jul 8, 2021 at 3:20 PM Rafael J. Wysocki <rjw@rjwysocki.net> wrote: > > > > On Thursday, July 8, 2021 2:39:49 PM CEST Rafael J. Wysocki wrote: > > > On Thu, Jul 8, 2021 at 2:18 PM Rafael J. Wysocki <rafael@kernel.org> wrote: > > > > > > > > On Wed, Jul 7, 2021 at 11:57 PM Bjorn Helgaas <helgaas@kernel.org> wrote: > > > > > > > > > > On Thu, Jun 17, 2021 at 03:36:53PM +0300, Mika Westerberg wrote: > > > > > > Some PCIe devices only support PME (Power Management Event) from D3cold. > > > > > > One example is ASMedia xHCI controller: > > > > > > > > > > > > 11:00.0 USB controller: ASMedia Technology Inc. ASM1042A USB 3.0 Host Controller (prog-if 30 [XHCI]) > > > > > > ... > > > > > > Capabilities: [78] Power Management version 3 > > > > > > Flags: PMEClk- DSI- D1- D2- AuxCurrent=55mA PME(D0-,D1-,D2-,D3hot-,D3cold+) > > > > > > Status: D0 NoSoftRst+ PME-Enable- DSel=0 DScale=0 PME- > > > > > > > > > > > > With such devices, if it has wake enabled, the kernel selects lowest > > > > > > possible power state to be D0 in pci_target_state(). This is problematic > > > > > > because it prevents the root port it is connected to enter low power > > > > > > state too which makes the system consume more energy than necessary. > > > > > > > > > > IIUC this is because the loop that checks which states support PME > > > > > starts with D3hot and doesn't even look at D3cold. > > > > > > > > That's because the device itself cannot be programmed into D3cold, so > > > > the target state cannot be D3cold for it. > > > > > > > > > > The problem in pci_target_state() is that it only accounts the "current" > > > > > > device state, so when the bridge above it (a root port for instance) is > > > > > > transitioned into D3hot the device transitions into D3cold. This is > > > > > > because when the root port is first transitioned into D3hot then the > > > > > > ACPI power resource is turned off which puts the PCIe link to L2/L3 (and > > > > > > the root port and the device are in D3cold). If the root port is kept in > > > > > > D3hot it still means that the device below it is still effectively in > > > > > > D3cold as no configuration messages pass through. Furthermore the > > > > > > implementation note of PCIe 5.0 sec 5.3.1.4 says that the device should > > > > > > expect to be transitioned into D3cold soon after its link transitions > > > > > > into L2/L3 Ready state. > > > > > > > > > > > > Taking the above into consideration, instead of forcing the device stay > > > > > > in D0 we modify pci_target_state() to return D3hot in this special case > > > > > > and make __pci_enable_wake() to enable PME too in this case. > > > > > > > > > > > > Reported-by: Utkarsh H Patel <utkarsh.h.patel@intel.com> > > > > > > Reported-by: Koba Ko <koba.ko@canonical.com> > > > > > > Suggested-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > > > > > > Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com> > > > > > > --- > > > > > > The previous version of the patch is here: > > > > > > > > > > > > https://lore.kernel.org/linux-pm/20210616150516.28242-1-mika.westerberg@linux.intel.com/ > > > > > > > > > > > > Changes from the previous version: > > > > > > > > > > > > * Dropped redundant test in pci_target_state(). > > > > > > > > > > > > drivers/pci/pci.c | 16 +++++++++++++++- > > > > > > 1 file changed, 15 insertions(+), 1 deletion(-) > > > > > > > > > > > > diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c > > > > > > index b717680377a9..043c5c304308 100644 > > > > > > --- a/drivers/pci/pci.c > > > > > > +++ b/drivers/pci/pci.c > > > > > > @@ -2485,7 +2485,13 @@ static int __pci_enable_wake(struct pci_dev *dev, pci_power_t state, bool enable > > > > > > if (enable) { > > > > > > int error; > > > > > > > > > > > > - if (pci_pme_capable(dev, state)) > > > > > > + /* > > > > > > + * Enable PME if device is capable from given state. > > > > > > + * Special case is device that can only generate PME > > > > > > + * from D3cold then we enable PME too. > > > > > > + */ > > > > > > + if (pci_pme_capable(dev, state) || > > > > > > + (state == PCI_D3hot && pci_pme_capable(dev, PCI_D3cold))) > > > > > > pci_pme_active(dev, true); > > > > > > else > > > > > > ret = 1; > > > > > > @@ -2595,6 +2601,14 @@ static pci_power_t pci_target_state(struct pci_dev *dev, bool wakeup) > > > > > > * PME#. > > > > > > */ > > > > > > if (dev->pme_support) { > > > > > > + /* > > > > > > + * Special case if device supports only PME from > > > > > > + * D3cold but not from D3hot we still return D3hot. > > > > > > + */ > > > > > > + if (target_state == PCI_D3hot && > > > > > > + (dev->pme_support & (1 << PCI_D3cold))) > > > > > > + return target_state; > > > > > > > > > > I've spent quite a bit of time trying to understand this, and I'm kind > > > > > of dragging my feet on it because I haven't been able to really > > > > > connect this with the specs. > > > > > > > > The specs aren't very clear in this area, though. > > > > > > > > The overall picture is that the device in question is connected to a > > > > port (a root port in this particular case) that can be programmed into > > > > D3cold via ACPI, but the endpoint itself can only be programmed into > > > > D3hot. However, if the port goes into D3cold, the endpoint also goes > > > > into D3cold (actually, my understanding of the specs is that even if > > > > the port goes into D3hot, the endpoint should still be assumed to go > > > > into D3cold). > > > > > > > > The power state of the endpoint is changed first and at the time this > > > > happens it is not known which power state the port is going to be > > > > programmed into. > > > > > > > > Now, the device is wake-capable (in general) and so we want it to be > > > > able to signal wakeup from the final power state. Because it only > > > > reports PME support in D0 and in D3cold, the kernel today leaves it in > > > > D0 which causes the port to stay in D0 too. Still, putting the device > > > > into D3hot allows the port to go into D3cold which in turn causes the > > > > device to go into D3cold and it can signal wakeup from that state. > > > > > > > > So there are two ways to get into a configuration from which the > > > > endpoint device can signal wakeup, either by leaving it and the port > > > > holding it both in D0, or by putting it into D3hot, so that the port > > > > can go into D3cold in which case the endpoint will end up in D3cold. > > > > > > > > The Mika's patch is aiming at enabling the second option. > > > > > > > > > It also seems unfortunate to have to add this special case in two places. > > > > > > > > That's because __pci_enable_wake() tries to be extra careful and only > > > > call pci_pme_active() if PME is known to be supported in the target > > > > power state, but that is not strictly necessary. It could just call > > > > pci_pme_active() unconditionally and return the > > > > platform_pci_set_wakeup() return value. > > > > > > > > I think I'll send a patch making this change. > > > > > > Actually, it needs to fail if PME cannot be signaled from the target > > > state and the device is not power-manageable by the platform. > > > > > > A better idea may be to make pci_pme_capable() also check if the > > > parent bridge can go into D3cold and return "true" if so and "state" > > > is D3hot while PME signaling from D3cold is supported. > > > > So below is my version of the $subject patch (untested). > > > > Please let me know what you think. > > I gave some more consideration to this and I was not able to convince > myself that putting the parent port into D3hot was sufficient for the > endpoint device connected to it to go into D3cold. Thanks, that was a sticking point for me, too. I've never been able to directly connect the parent port's D0-D3cold power state to the main power state for a downstream device. > However, the PCI PM spec v1.2 clearly mandates that putting a bridge > into D3cold will cause power to be removed from the entire bus > segment below it, which should apply to PCIe devices by extension. Are you referring to Table 6-1, where it says "No PCI transactions; no clock; no Vcc" on the secondary bus if the bridge is in D3cold? I'm not really clear on how a bridge's power state affects main power for downstream devices. What about optical links where the power distribution is more separate from the communication path? > So I'm going to submit a new version of the patch below in which > pci_pme_capable() will check whether or not the parent will go into > D3cold in addition to all of the other checks. > > Stay tuned! Bjorn
On Thu, Jul 15, 2021 at 6:14 PM Bjorn Helgaas <helgaas@kernel.org> wrote: > > On Mon, Jul 12, 2021 at 05:22:58PM +0200, Rafael J. Wysocki wrote: > > On Thu, Jul 8, 2021 at 3:20 PM Rafael J. Wysocki <rjw@rjwysocki.net> wrote: > > > > > > On Thursday, July 8, 2021 2:39:49 PM CEST Rafael J. Wysocki wrote: > > > > On Thu, Jul 8, 2021 at 2:18 PM Rafael J. Wysocki <rafael@kernel.org> wrote: > > > > > > > > > > On Wed, Jul 7, 2021 at 11:57 PM Bjorn Helgaas <helgaas@kernel.org> wrote: > > > > > > > > > > > > On Thu, Jun 17, 2021 at 03:36:53PM +0300, Mika Westerberg wrote: > > > > > > > Some PCIe devices only support PME (Power Management Event) from D3cold. > > > > > > > One example is ASMedia xHCI controller: > > > > > > > > > > > > > > 11:00.0 USB controller: ASMedia Technology Inc. ASM1042A USB 3.0 Host Controller (prog-if 30 [XHCI]) > > > > > > > ... > > > > > > > Capabilities: [78] Power Management version 3 > > > > > > > Flags: PMEClk- DSI- D1- D2- AuxCurrent=55mA PME(D0-,D1-,D2-,D3hot-,D3cold+) > > > > > > > Status: D0 NoSoftRst+ PME-Enable- DSel=0 DScale=0 PME- > > > > > > > > > > > > > > With such devices, if it has wake enabled, the kernel selects lowest > > > > > > > possible power state to be D0 in pci_target_state(). This is problematic > > > > > > > because it prevents the root port it is connected to enter low power > > > > > > > state too which makes the system consume more energy than necessary. > > > > > > > > > > > > IIUC this is because the loop that checks which states support PME > > > > > > starts with D3hot and doesn't even look at D3cold. > > > > > > > > > > That's because the device itself cannot be programmed into D3cold, so > > > > > the target state cannot be D3cold for it. > > > > > > > > > > > > The problem in pci_target_state() is that it only accounts the "current" > > > > > > > device state, so when the bridge above it (a root port for instance) is > > > > > > > transitioned into D3hot the device transitions into D3cold. This is > > > > > > > because when the root port is first transitioned into D3hot then the > > > > > > > ACPI power resource is turned off which puts the PCIe link to L2/L3 (and > > > > > > > the root port and the device are in D3cold). If the root port is kept in > > > > > > > D3hot it still means that the device below it is still effectively in > > > > > > > D3cold as no configuration messages pass through. Furthermore the > > > > > > > implementation note of PCIe 5.0 sec 5.3.1.4 says that the device should > > > > > > > expect to be transitioned into D3cold soon after its link transitions > > > > > > > into L2/L3 Ready state. > > > > > > > > > > > > > > Taking the above into consideration, instead of forcing the device stay > > > > > > > in D0 we modify pci_target_state() to return D3hot in this special case > > > > > > > and make __pci_enable_wake() to enable PME too in this case. > > > > > > > > > > > > > > Reported-by: Utkarsh H Patel <utkarsh.h.patel@intel.com> > > > > > > > Reported-by: Koba Ko <koba.ko@canonical.com> > > > > > > > Suggested-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > > > > > > > Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com> > > > > > > > --- > > > > > > > The previous version of the patch is here: > > > > > > > > > > > > > > https://lore.kernel.org/linux-pm/20210616150516.28242-1-mika.westerberg@linux.intel.com/ > > > > > > > > > > > > > > Changes from the previous version: > > > > > > > > > > > > > > * Dropped redundant test in pci_target_state(). > > > > > > > > > > > > > > drivers/pci/pci.c | 16 +++++++++++++++- > > > > > > > 1 file changed, 15 insertions(+), 1 deletion(-) > > > > > > > > > > > > > > diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c > > > > > > > index b717680377a9..043c5c304308 100644 > > > > > > > --- a/drivers/pci/pci.c > > > > > > > +++ b/drivers/pci/pci.c > > > > > > > @@ -2485,7 +2485,13 @@ static int __pci_enable_wake(struct pci_dev *dev, pci_power_t state, bool enable > > > > > > > if (enable) { > > > > > > > int error; > > > > > > > > > > > > > > - if (pci_pme_capable(dev, state)) > > > > > > > + /* > > > > > > > + * Enable PME if device is capable from given state. > > > > > > > + * Special case is device that can only generate PME > > > > > > > + * from D3cold then we enable PME too. > > > > > > > + */ > > > > > > > + if (pci_pme_capable(dev, state) || > > > > > > > + (state == PCI_D3hot && pci_pme_capable(dev, PCI_D3cold))) > > > > > > > pci_pme_active(dev, true); > > > > > > > else > > > > > > > ret = 1; > > > > > > > @@ -2595,6 +2601,14 @@ static pci_power_t pci_target_state(struct pci_dev *dev, bool wakeup) > > > > > > > * PME#. > > > > > > > */ > > > > > > > if (dev->pme_support) { > > > > > > > + /* > > > > > > > + * Special case if device supports only PME from > > > > > > > + * D3cold but not from D3hot we still return D3hot. > > > > > > > + */ > > > > > > > + if (target_state == PCI_D3hot && > > > > > > > + (dev->pme_support & (1 << PCI_D3cold))) > > > > > > > + return target_state; > > > > > > > > > > > > I've spent quite a bit of time trying to understand this, and I'm kind > > > > > > of dragging my feet on it because I haven't been able to really > > > > > > connect this with the specs. > > > > > > > > > > The specs aren't very clear in this area, though. > > > > > > > > > > The overall picture is that the device in question is connected to a > > > > > port (a root port in this particular case) that can be programmed into > > > > > D3cold via ACPI, but the endpoint itself can only be programmed into > > > > > D3hot. However, if the port goes into D3cold, the endpoint also goes > > > > > into D3cold (actually, my understanding of the specs is that even if > > > > > the port goes into D3hot, the endpoint should still be assumed to go > > > > > into D3cold). > > > > > > > > > > The power state of the endpoint is changed first and at the time this > > > > > happens it is not known which power state the port is going to be > > > > > programmed into. > > > > > > > > > > Now, the device is wake-capable (in general) and so we want it to be > > > > > able to signal wakeup from the final power state. Because it only > > > > > reports PME support in D0 and in D3cold, the kernel today leaves it in > > > > > D0 which causes the port to stay in D0 too. Still, putting the device > > > > > into D3hot allows the port to go into D3cold which in turn causes the > > > > > device to go into D3cold and it can signal wakeup from that state. > > > > > > > > > > So there are two ways to get into a configuration from which the > > > > > endpoint device can signal wakeup, either by leaving it and the port > > > > > holding it both in D0, or by putting it into D3hot, so that the port > > > > > can go into D3cold in which case the endpoint will end up in D3cold. > > > > > > > > > > The Mika's patch is aiming at enabling the second option. > > > > > > > > > > > It also seems unfortunate to have to add this special case in two places. > > > > > > > > > > That's because __pci_enable_wake() tries to be extra careful and only > > > > > call pci_pme_active() if PME is known to be supported in the target > > > > > power state, but that is not strictly necessary. It could just call > > > > > pci_pme_active() unconditionally and return the > > > > > platform_pci_set_wakeup() return value. > > > > > > > > > > I think I'll send a patch making this change. > > > > > > > > Actually, it needs to fail if PME cannot be signaled from the target > > > > state and the device is not power-manageable by the platform. > > > > > > > > A better idea may be to make pci_pme_capable() also check if the > > > > parent bridge can go into D3cold and return "true" if so and "state" > > > > is D3hot while PME signaling from D3cold is supported. > > > > > > So below is my version of the $subject patch (untested). > > > > > > Please let me know what you think. > > > > I gave some more consideration to this and I was not able to convince > > myself that putting the parent port into D3hot was sufficient for the > > endpoint device connected to it to go into D3cold. > > Thanks, that was a sticking point for me, too. I've never been able > to directly connect the parent port's D0-D3cold power state to the > main power state for a downstream device. > > > However, the PCI PM spec v1.2 clearly mandates that putting a bridge > > into D3cold will cause power to be removed from the entire bus > > segment below it, which should apply to PCIe devices by extension. > > Are you referring to Table 6-1, where it says "No PCI transactions; no > clock; no Vcc" on the secondary bus if the bridge is in D3cold? Yes, I am. > I'm not really clear on how a bridge's power state affects main power > for downstream devices. What about optical links where the power > distribution is more separate from the communication path? My understanding is that the PCIe downstream port going into D3cold must shut down the link to the downstream component connected to it which means that, as far as PME signaling goes, the downstream component ends up in the D3cold-equivalent situation, even if it has its own power supply.
diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c index b717680377a9..043c5c304308 100644 --- a/drivers/pci/pci.c +++ b/drivers/pci/pci.c @@ -2485,7 +2485,13 @@ static int __pci_enable_wake(struct pci_dev *dev, pci_power_t state, bool enable if (enable) { int error; - if (pci_pme_capable(dev, state)) + /* + * Enable PME if device is capable from given state. + * Special case is device that can only generate PME + * from D3cold then we enable PME too. + */ + if (pci_pme_capable(dev, state) || + (state == PCI_D3hot && pci_pme_capable(dev, PCI_D3cold))) pci_pme_active(dev, true); else ret = 1; @@ -2595,6 +2601,14 @@ static pci_power_t pci_target_state(struct pci_dev *dev, bool wakeup) * PME#. */ if (dev->pme_support) { + /* + * Special case if device supports only PME from + * D3cold but not from D3hot we still return D3hot. + */ + if (target_state == PCI_D3hot && + (dev->pme_support & (1 << PCI_D3cold))) + return target_state; + while (target_state && !(dev->pme_support & (1 << target_state))) target_state--;
Some PCIe devices only support PME (Power Management Event) from D3cold. One example is ASMedia xHCI controller: 11:00.0 USB controller: ASMedia Technology Inc. ASM1042A USB 3.0 Host Controller (prog-if 30 [XHCI]) ... Capabilities: [78] Power Management version 3 Flags: PMEClk- DSI- D1- D2- AuxCurrent=55mA PME(D0-,D1-,D2-,D3hot-,D3cold+) Status: D0 NoSoftRst+ PME-Enable- DSel=0 DScale=0 PME- With such devices, if it has wake enabled, the kernel selects lowest possible power state to be D0 in pci_target_state(). This is problematic because it prevents the root port it is connected to enter low power state too which makes the system consume more energy than necessary. The problem in pci_target_state() is that it only accounts the "current" device state, so when the bridge above it (a root port for instance) is transitioned into D3hot the device transitions into D3cold. This is because when the root port is first transitioned into D3hot then the ACPI power resource is turned off which puts the PCIe link to L2/L3 (and the root port and the device are in D3cold). If the root port is kept in D3hot it still means that the device below it is still effectively in D3cold as no configuration messages pass through. Furthermore the implementation note of PCIe 5.0 sec 5.3.1.4 says that the device should expect to be transitioned into D3cold soon after its link transitions into L2/L3 Ready state. Taking the above into consideration, instead of forcing the device stay in D0 we modify pci_target_state() to return D3hot in this special case and make __pci_enable_wake() to enable PME too in this case. Reported-by: Utkarsh H Patel <utkarsh.h.patel@intel.com> Reported-by: Koba Ko <koba.ko@canonical.com> Suggested-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com> --- The previous version of the patch is here: https://lore.kernel.org/linux-pm/20210616150516.28242-1-mika.westerberg@linux.intel.com/ Changes from the previous version: * Dropped redundant test in pci_target_state(). drivers/pci/pci.c | 16 +++++++++++++++- 1 file changed, 15 insertions(+), 1 deletion(-)