Message ID | 20230915023354.939-3-mario.limonciello@amd.com (mailing list archive) |
---|---|
State | Superseded |
Headers | show |
Series | Add quirk for PCIe root port on AMD systems | expand |
On Thu, Sep 14, 2023 at 09:33:54PM -0500, Mario Limonciello wrote: > +static bool child_has_amd_usb4(struct pci_dev *pdev) > +{ > + struct pci_dev *child = NULL; > + > + while ((child = pci_get_class(PCI_CLASS_SERIAL_USB_USB4, child))) { > + if (child->vendor != PCI_VENDOR_ID_AMD) > + continue; > + if (pcie_find_root_port(child) != pdev) > + continue; > + return true; > + } > + > + return false; > +} What's the purpose of the pcie_find_root_port() check? PCI is a hierarchy, not a graph, so a device cannot have any other Root Port but the one below which you're searching. If the purpose is to check that the port is a Root Port (if the PCI IDs you're using in the DECLARE_PCI_FIXUP_* clauses match non-Root Ports), check for pci_pcie_type(pdev) == PCI_EXP_TYPE_ROOT_PORT. (No need to check for that in every loop iteration obviously, just check once in the fixup.) Thanks, Lukas
On 9/15/2023 02:08, Lukas Wunner wrote: > On Thu, Sep 14, 2023 at 09:33:54PM -0500, Mario Limonciello wrote: >> +static bool child_has_amd_usb4(struct pci_dev *pdev) >> +{ >> + struct pci_dev *child = NULL; >> + >> + while ((child = pci_get_class(PCI_CLASS_SERIAL_USB_USB4, child))) { >> + if (child->vendor != PCI_VENDOR_ID_AMD) >> + continue; >> + if (pcie_find_root_port(child) != pdev) >> + continue; >> + return true; >> + } >> + >> + return false; >> +} > > What's the purpose of the pcie_find_root_port() check? PCI is a hierarchy, > not a graph, so a device cannot have any other Root Port but the one below > which you're searching. > > If the purpose is to check that the port is a Root Port (if the PCI IDs > you're using in the DECLARE_PCI_FIXUP_* clauses match non-Root Ports), > check for pci_pcie_type(pdev) == PCI_EXP_TYPE_ROOT_PORT. (No need to > check for that in every loop iteration obviously, just check once in > the fixup.) > > Thanks, > > Lukas The reason to look for it the way that I did was that there are multiple root ports with the exact same PCI ID. The problem only occurs on the root port that happens to have an AMD USB4 controller connected.
On Fri, Sep 15, 2023 at 07:04:11AM -0500, Mario Limonciello wrote: > On 9/15/2023 02:08, Lukas Wunner wrote: > > On Thu, Sep 14, 2023 at 09:33:54PM -0500, Mario Limonciello wrote: > > > +static bool child_has_amd_usb4(struct pci_dev *pdev) > > > +{ > > > + struct pci_dev *child = NULL; > > > + > > > + while ((child = pci_get_class(PCI_CLASS_SERIAL_USB_USB4, child))) { > > > + if (child->vendor != PCI_VENDOR_ID_AMD) > > > + continue; > > > + if (pcie_find_root_port(child) != pdev) > > > + continue; > > > + return true; > > > + } > > > + > > > + return false; > > > +} > > > > What's the purpose of the pcie_find_root_port() check? PCI is a hierarchy, > > not a graph, so a device cannot have any other Root Port but the one below > > which you're searching. > > > > If the purpose is to check that the port is a Root Port (if the PCI IDs > > you're using in the DECLARE_PCI_FIXUP_* clauses match non-Root Ports), > > check for pci_pcie_type(pdev) == PCI_EXP_TYPE_ROOT_PORT. (No need to > > check for that in every loop iteration obviously, just check once in > > the fixup.) > > > > Thanks, > > > > Lukas > > The reason to look for it the way that I did was that there are multiple > root ports with the exact same PCI ID. > > The problem only occurs on the root port that happens to have an AMD USB4 > controller connected. Yes but what's the purpose of the pcie_find_root_port(child) check quoted above? Thanks, Lukas
On 9/15/2023 23:48, Lukas Wunner wrote: > On Fri, Sep 15, 2023 at 07:04:11AM -0500, Mario Limonciello wrote: >> On 9/15/2023 02:08, Lukas Wunner wrote: >>> On Thu, Sep 14, 2023 at 09:33:54PM -0500, Mario Limonciello wrote: >>>> +static bool child_has_amd_usb4(struct pci_dev *pdev) >>>> +{ >>>> + struct pci_dev *child = NULL; >>>> + >>>> + while ((child = pci_get_class(PCI_CLASS_SERIAL_USB_USB4, child))) { >>>> + if (child->vendor != PCI_VENDOR_ID_AMD) >>>> + continue; >>>> + if (pcie_find_root_port(child) != pdev) >>>> + continue; >>>> + return true; >>>> + } >>>> + >>>> + return false; >>>> +} >>> >>> What's the purpose of the pcie_find_root_port() check? PCI is a hierarchy, >>> not a graph, so a device cannot have any other Root Port but the one below >>> which you're searching. >>> >>> If the purpose is to check that the port is a Root Port (if the PCI IDs >>> you're using in the DECLARE_PCI_FIXUP_* clauses match non-Root Ports), >>> check for pci_pcie_type(pdev) == PCI_EXP_TYPE_ROOT_PORT. (No need to >>> check for that in every loop iteration obviously, just check once in >>> the fixup.) >>> >>> Thanks, >>> >>> Lukas >> >> The reason to look for it the way that I did was that there are multiple >> root ports with the exact same PCI ID. >> >> The problem only occurs on the root port that happens to have an AMD USB4 >> controller connected. > > Yes but what's the purpose of the pcie_find_root_port(child) check > quoted above? > > Thanks, > > Lukas You're right that if you look at this system alone that the check isn't strictly necessary. It's to future proof the quirk. If a discrete USB4 controller was connected to the system it would be connected to a different root port (the one that is used for PCI tunneling). AMD doesn't have any of these devices, but if some day one was created it could trip this codepath. If you feel it's better to remove the check unless such a device is created sure I can drop it.
On Sat, Sep 16, 2023 at 08:09:19AM -0500, Mario Limonciello wrote: > On 9/15/2023 23:48, Lukas Wunner wrote: > > On Fri, Sep 15, 2023 at 07:04:11AM -0500, Mario Limonciello wrote: > > > On 9/15/2023 02:08, Lukas Wunner wrote: > > > > On Thu, Sep 14, 2023 at 09:33:54PM -0500, Mario Limonciello wrote: > > > > > +static bool child_has_amd_usb4(struct pci_dev *pdev) > > > > > +{ > > > > > + struct pci_dev *child = NULL; > > > > > + > > > > > + while ((child = pci_get_class(PCI_CLASS_SERIAL_USB_USB4, child))) { > > > > > + if (child->vendor != PCI_VENDOR_ID_AMD) > > > > > + continue; > > > > > + if (pcie_find_root_port(child) != pdev) > > > > > + continue; > > > > > + return true; > > > > > + } > > > > > + > > > > > + return false; > > > > > +} > > > > > > > > What's the purpose of the pcie_find_root_port() check? PCI is a hierarchy, > > > > not a graph, so a device cannot have any other Root Port but the one below > > > > which you're searching. > > > > > > > > If the purpose is to check that the port is a Root Port (if the PCI IDs > > > > you're using in the DECLARE_PCI_FIXUP_* clauses match non-Root Ports), > > > > check for pci_pcie_type(pdev) == PCI_EXP_TYPE_ROOT_PORT. (No need to > > > > check for that in every loop iteration obviously, just check once in > > > > the fixup.) > > > > > > > > Thanks, > > > > > > > > Lukas > > > > > > The reason to look for it the way that I did was that there are multiple > > > root ports with the exact same PCI ID. > > > > > > The problem only occurs on the root port that happens to have an AMD USB4 > > > controller connected. > > > > Yes but what's the purpose of the pcie_find_root_port(child) check > > quoted above? > > You're right that if you look at this system alone that the check isn't > strictly necessary. It's to future proof the quirk. If a discrete USB4 > controller was connected to the system it would be connected to a different > root port (the one that is used for PCI tunneling). > > AMD doesn't have any of these devices, but if some day one was created it > could trip this codepath. > > If you feel it's better to remove the check unless such a device is created > sure I can drop it. PCIe ports used for Thunderbolt tunneling are Downstream Ports or Upstream Ports (depending on which of the two ends of the tunnel you're looking at). The "pcie_find_root_port(child) != pdev" check is always false: You're searching for a USB controller below a Root Port and check whether the Root Port in the USB controller's ancestry is the Root Port below which you're searching. That's a tautology. I'm guessing what you really mean is: if (pci_upstream_bridge(child)) != pdev) continue; Thanks, Lukas
On 9/16/2023 08:36, Lukas Wunner wrote: > On Sat, Sep 16, 2023 at 08:09:19AM -0500, Mario Limonciello wrote: >> On 9/15/2023 23:48, Lukas Wunner wrote: >>> On Fri, Sep 15, 2023 at 07:04:11AM -0500, Mario Limonciello wrote: >>>> On 9/15/2023 02:08, Lukas Wunner wrote: >>>>> On Thu, Sep 14, 2023 at 09:33:54PM -0500, Mario Limonciello wrote: >>>>>> +static bool child_has_amd_usb4(struct pci_dev *pdev) >>>>>> +{ >>>>>> + struct pci_dev *child = NULL; >>>>>> + >>>>>> + while ((child = pci_get_class(PCI_CLASS_SERIAL_USB_USB4, child))) { >>>>>> + if (child->vendor != PCI_VENDOR_ID_AMD) >>>>>> + continue; >>>>>> + if (pcie_find_root_port(child) != pdev) >>>>>> + continue; >>>>>> + return true; >>>>>> + } >>>>>> + >>>>>> + return false; >>>>>> +} >>>>> >>>>> What's the purpose of the pcie_find_root_port() check? PCI is a hierarchy, >>>>> not a graph, so a device cannot have any other Root Port but the one below >>>>> which you're searching. >>>>> >>>>> If the purpose is to check that the port is a Root Port (if the PCI IDs >>>>> you're using in the DECLARE_PCI_FIXUP_* clauses match non-Root Ports), >>>>> check for pci_pcie_type(pdev) == PCI_EXP_TYPE_ROOT_PORT. (No need to >>>>> check for that in every loop iteration obviously, just check once in >>>>> the fixup.) >>>>> >>>>> Thanks, >>>>> >>>>> Lukas >>>> >>>> The reason to look for it the way that I did was that there are multiple >>>> root ports with the exact same PCI ID. >>>> >>>> The problem only occurs on the root port that happens to have an AMD USB4 >>>> controller connected. >>> >>> Yes but what's the purpose of the pcie_find_root_port(child) check >>> quoted above? >> >> You're right that if you look at this system alone that the check isn't >> strictly necessary. It's to future proof the quirk. If a discrete USB4 >> controller was connected to the system it would be connected to a different >> root port (the one that is used for PCI tunneling). >> >> AMD doesn't have any of these devices, but if some day one was created it >> could trip this codepath. >> >> If you feel it's better to remove the check unless such a device is created >> sure I can drop it. > > PCIe ports used for Thunderbolt tunneling are Downstream Ports or > Upstream Ports (depending on which of the two ends of the tunnel > you're looking at). Let me explain the topology from an actual system with PCIe tunneling active and a dock connected downstream. -[0000:00]-+-00.0 +-04.1-[33-62]----00.0-[34-36]--+-02.0-[35]----00.0 | \-04.0-[36]-- +-08.3-[64]--+-00.0 | +-00.3 | +-00.4 | +-00.5 | \-00.6 This is the root port that tunneled devices will connect to: $ sudo lspci -s 0000:00:04.1 -v | grep "Capabilities: \[58\]" Capabilities: [58] Express Root Port (Slot+), MSI 00 This is the root port that the USB4 controllers are connected to that is problematic: $ sudo lspci -s 0000:00:08.3 -v | grep "Capabilities: \[58\]" Capabilities: [58] Express Root Port (Slot-), MSI 00 Here's a downstream connected dock from tunneling: 33:00.0 PCI bridge [0604]: Intel Corporation JHL7540 Thunderbolt 3 Bridge [Titan Ridge DD 2018] [8086:15ef] (rev 06) 34:02.0 PCI bridge [0604]: Intel Corporation JHL7540 Thunderbolt 3 Bridge [Titan Ridge DD 2018] [8086:15ef] (rev 06) 34:04.0 PCI bridge [0604]: Intel Corporation JHL7540 Thunderbolt 3 Bridge [Titan Ridge DD 2018] [8086:15ef] (rev 06) Here's the USB4 controllers connected to 08.3: 64:00.0 USB controller [0c03]: Advanced Micro Devices, Inc. [AMD] Device [1022:161f] 64:00.3 USB controller [0c03]: Advanced Micro Devices, Inc. [AMD] Device [1022:15d6] 64:00.4 USB controller [0c03]: Advanced Micro Devices, Inc. [AMD] Device [1022:15d7] 64:00.5 USB controller [0c03]: Advanced Micro Devices, Inc. [AMD] Device [1022:162e] 64:00.6 USB controller [0c03]: Advanced Micro Devices, Inc. [AMD] Device [1022:162f] If an AMD USB4 controller happened to be connected, it would have shown up connected to 04.1 as it's root port. > > The "pcie_find_root_port(child) != pdev" check is always false: > > You're searching for a USB controller below a Root Port and > check whether the Root Port in the USB controller's ancestry > is the Root Port below which you're searching. That's a > tautology. The search doesn't start at pdev, it starts at NULL. struct pci_dev *child = NULL; > > I'm guessing what you really mean is: > > if (pci_upstream_bridge(child)) != pdev) > continue; > That's exactly what I had before and Rafael suggested [1] to change it to pcie_find_root_port(). [1] https://lore.kernel.org/linux-usb/20230913040832.114610-1-mario.limonciello@amd.com/T/#m66acb79a13d314b5e868993b1611266a968cf064 > Thanks, > > Lukas
On Sat, Sep 16, 2023 at 03:36:50PM +0200, Lukas Wunner wrote: > On Sat, Sep 16, 2023 at 08:09:19AM -0500, Mario Limonciello wrote: > > On 9/15/2023 23:48, Lukas Wunner wrote: > > > On Fri, Sep 15, 2023 at 07:04:11AM -0500, Mario Limonciello wrote: > > > > On 9/15/2023 02:08, Lukas Wunner wrote: > > > > > On Thu, Sep 14, 2023 at 09:33:54PM -0500, Mario Limonciello wrote: > > > > > > +static bool child_has_amd_usb4(struct pci_dev *pdev) > > > > > > +{ > > > > > > + struct pci_dev *child = NULL; > > > > > > + > > > > > > + while ((child = pci_get_class(PCI_CLASS_SERIAL_USB_USB4, child))) { > > > > > > + if (child->vendor != PCI_VENDOR_ID_AMD) > > > > > > + continue; > > > > > > + if (pcie_find_root_port(child) != pdev) > > > > > > + continue; > > > > > > + return true; > > > > > > + } > > > > > > + > > > > > > + return false; > > > > > > +} > ... > The "pcie_find_root_port(child) != pdev" check is always false: If we were using pci_walk_bus() and only looking at devices below pdev, I would agree, but since we're using pci_get_class(), which searches all PCI devices in the system, I'm confused about why it would always be false. I don't really see the point of checking for USB4, because the commit log doesn't say anything about why this would be specific to USB4. I know Mario has mentioned something about how "internal interrupt routing works with the USB4 controller connected to this root port," but I don't understand what that means. Is the USB4 controller integrated into the Root Port? Or is this interrupt routed via some non-PCIe mechanism? If the USB4 controller is connected via standard PCIe, I don't see why the issue sould be specific to USB4. I could believe that BIOS configures the Root Port differently based on whether the downstream device is USB4, but I haven't heard anything about that. Bjorn
On Thu, Sep 14, 2023 at 09:33:54PM -0500, Mario Limonciello wrote: > Iain reports that USB devices can't be used to wake a Lenovo Z13 > from suspend. This problem occurs because the PCIe root port has been put > into D3hot and AMD's platform can't handle USB devices waking the platform > from a hardware sleep state in this case. The platform is put into > a hardware sleep state by the actions of the amd-pmc driver. > > Although the issue is initially reported on a single model it actually > affects all Yellow Carp (Rembrandt) and Pink Sardine (Phoenix) SoCs. > This problem only occurs on Linux specifically when attempting to > wake the platform from a hardware sleep state. > Comparing the behavior on Windows and Linux, Windows doesn't put > the root ports into D3 at this time. > > Linux decides the target state to put the device into at suspend by > this policy: > 1. If platform_pci_power_manageable(): > Use platform_pci_choose_state() > 2. If the device is armed for wakeup: > Select the deepest D-state that supports a PME. > 3. Else: > Use D3hot. > > Devices are considered power manageable by the platform when they have > one or more objects described in the table in section 7.3 of the ACPI 6.5 > specification [1]. In this case the root ports are not power manageable. > > If devices are not considered power manageable; specs are ambiguous as > to what should happen. In this situation Windows 11 puts PCIe ports > in D0 ostensibly due the policy from the "uPEP driver" which is a > complimentary driver the Linux "amd-pmc" driver. > > Linux chooses to allow D3 for these root ports due to the policy > introduced by commit 9d26d3a8f1b0 ("PCI: Put PCIe ports into D3 during > suspend"). > > The Windows uPEP driver expresses the desired state that should be > selected for suspend but Linux doesn't, so introduce a quirk for the > problematic root ports. > > The quirk removes PME support for D3hot and D3cold at suspend time if the > system will be using s2idle. When the port is configured for wakeup this > will prevent these states from being selected in pci_target_state(). > > After the system is resumes the PME support is re-read from the PM > capabilities register to allow opportunistic power savings at runtime by > letting the root port go into D3hot or D3cold. There's a lot of text here, but I think the essential thing is: These Root Ports advertise D3hot and D3cold in the PME_Support register, but PMEs do not work in those states when the amd-pmc driver has put the platform in a sleep state. > Cc: stable@vger.kernel.org > Link: https://learn.microsoft.com/en-us/windows-hardware/design/device-experiences/platform-design-for-modern-standby#low-power-core-silicon-cpu-soc-dram [1] > Fixes: 9d26d3a8f1b0 ("PCI: Put PCIe ports into D3 during suspend") > Reported-by: Iain Lane <iain@orangesquash.org.uk> > Closes: https://forums.lenovo.com/t5/Ubuntu/Z13-can-t-resume-from-suspend-with-external-USB-keyboard/m-p/5217121 > Signed-off-by: Mario Limonciello <mario.limonciello@amd.com> > --- > drivers/pci/quirks.c | 61 ++++++++++++++++++++++++++++++++++++++++++++ > 1 file changed, 61 insertions(+) > > diff --git a/drivers/pci/quirks.c b/drivers/pci/quirks.c > index eeec1d6f9023..ebc0afbc814e 100644 > --- a/drivers/pci/quirks.c > +++ b/drivers/pci/quirks.c > @@ -6188,3 +6188,64 @@ DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x9a31, dpc_log_size); > DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_XILINX, 0x5020, of_pci_make_dev_node); > DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_XILINX, 0x5021, of_pci_make_dev_node); > DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_REDHAT, 0x0005, of_pci_make_dev_node); > + > +/* > + * When AMD PCIe root ports with AMD USB4 controllers attached to them are put > + * into D3hot or D3cold downstream USB devices may fail to wakeup the system. > + * This manifests as a missing wakeup interrupt. > + * > + * Prevent the associated root port from using PME to wake from D3hot or > + * D3cold power states during s2idle. > + * This will effectively put the root port into D0 power state over s2idle. > + */ > +static bool child_has_amd_usb4(struct pci_dev *pdev) > +{ > + struct pci_dev *child = NULL; > + > + while ((child = pci_get_class(PCI_CLASS_SERIAL_USB_USB4, child))) { > + if (child->vendor != PCI_VENDOR_ID_AMD) > + continue; > + if (pcie_find_root_port(child) != pdev) > + continue; > + return true; > + } > + > + return false; > +} > + > +static void quirk_reenable_pme(struct pci_dev *dev) > +{ > + u16 pmc; > + > + if (!dev->pm_cap) > + return; > + > + if (!child_has_amd_usb4(dev)) > + return; > + > + pci_read_config_word(dev, dev->pm_cap + PCI_PM_PMC, &pmc); > + pmc &= PCI_PM_CAP_PME_MASK; > + dev->pme_support = pmc >> PCI_PM_CAP_PME_SHIFT; > +} > + > +static void quirk_disable_pme_suspend(struct pci_dev *dev) > +{ > + int mask; > + > + if (pm_suspend_via_firmware()) > + return; There's always something more to confuse me. Why does pm_suspend_via_firmware() matter? I can sort of see that Linux platform power management, which uses the PMC, is not the same as platform firmware being invoked at the end of a system-wide power management transition to a sleep state. I guess this must have something to do with acpi_suspend_begin() and acpi_hibernation_begin() (the callers of pm_set_suspend_via_firmware())? > + if (!child_has_amd_usb4(dev)) > + return; > + > + mask = (PCI_PM_CAP_PME_D3hot|PCI_PM_CAP_PME_D3cold) >> PCI_PM_CAP_PME_SHIFT; > + if (!(dev->pme_support & mask)) > + return; > + dev->pme_support &= ~mask; > + dev_info_once(&dev->dev, "quirk: disabling PME from D3hot and D3cold at suspend\n"); > +} > + > +DECLARE_PCI_FIXUP_SUSPEND(PCI_VENDOR_ID_AMD, 0x14b9, quirk_disable_pme_suspend); > +DECLARE_PCI_FIXUP_RESUME(PCI_VENDOR_ID_AMD, 0x14b9, quirk_reenable_pme); > +DECLARE_PCI_FIXUP_SUSPEND(PCI_VENDOR_ID_AMD, 0x14eb, quirk_disable_pme_suspend); > +DECLARE_PCI_FIXUP_RESUME(PCI_VENDOR_ID_AMD, 0x14eb, quirk_reenable_pme); > -- > 2.34.1 >
On 9/17/2023 16:40, Bjorn Helgaas wrote: >> The "pcie_find_root_port(child) != pdev" check is always false: > > If we were using pci_walk_bus() and only looking at devices below > pdev, I would agree, but since we're using pci_get_class(), which > searches all PCI devices in the system, I'm confused about why it > would always be false. > > I don't really see the point of checking for USB4, because the commit > log doesn't say anything about why this would be specific to USB4. > > I know Mario has mentioned something about how "internal interrupt > routing works with the USB4 controller connected to this root port," > but I don't understand what that means. > > Is the USB4 controller integrated into the Root Port? Or is this > interrupt routed via some non-PCIe mechanism? If the USB4 controller > is connected via standard PCIe, I don't see why the issue sould be > specific to USB4. It's the latter. When the PMC has the SoC in hardware sleep the interrupt routing works differently. That's where this bug stems from. > > I could believe that BIOS configures the Root Port differently based > on whether the downstream device is USB4, but I haven't heard anything > about that. > > Bjorn Nothing along these lines.
On 9/17/2023 16:56, Bjorn Helgaas wrote: > On Thu, Sep 14, 2023 at 09:33:54PM -0500, Mario Limonciello wrote: >> Iain reports that USB devices can't be used to wake a Lenovo Z13 >> from suspend. This problem occurs because the PCIe root port has been put >> into D3hot and AMD's platform can't handle USB devices waking the platform >> from a hardware sleep state in this case. The platform is put into >> a hardware sleep state by the actions of the amd-pmc driver. >> >> Although the issue is initially reported on a single model it actually >> affects all Yellow Carp (Rembrandt) and Pink Sardine (Phoenix) SoCs. >> This problem only occurs on Linux specifically when attempting to >> wake the platform from a hardware sleep state. >> Comparing the behavior on Windows and Linux, Windows doesn't put >> the root ports into D3 at this time. >> >> Linux decides the target state to put the device into at suspend by >> this policy: >> 1. If platform_pci_power_manageable(): >> Use platform_pci_choose_state() >> 2. If the device is armed for wakeup: >> Select the deepest D-state that supports a PME. >> 3. Else: >> Use D3hot. >> >> Devices are considered power manageable by the platform when they have >> one or more objects described in the table in section 7.3 of the ACPI 6.5 >> specification [1]. In this case the root ports are not power manageable. >> >> If devices are not considered power manageable; specs are ambiguous as >> to what should happen. In this situation Windows 11 puts PCIe ports >> in D0 ostensibly due the policy from the "uPEP driver" which is a >> complimentary driver the Linux "amd-pmc" driver. >> >> Linux chooses to allow D3 for these root ports due to the policy >> introduced by commit 9d26d3a8f1b0 ("PCI: Put PCIe ports into D3 during >> suspend"). >> >> The Windows uPEP driver expresses the desired state that should be >> selected for suspend but Linux doesn't, so introduce a quirk for the >> problematic root ports. >> >> The quirk removes PME support for D3hot and D3cold at suspend time if the >> system will be using s2idle. When the port is configured for wakeup this >> will prevent these states from being selected in pci_target_state(). >> >> After the system is resumes the PME support is re-read from the PM >> capabilities register to allow opportunistic power savings at runtime by >> letting the root port go into D3hot or D3cold. > > There's a lot of text here, but I think the essential thing is: > > These Root Ports advertise D3hot and D3cold in the PME_Support > register, but PMEs do not work in those states when the amd-pmc > driver has put the platform in a sleep state. > It's specific to the PMEs for root ports with USB4 controller connected, but yes otherwise correct. I've confirmed that XHCI controllers connected to other root ports work fine. >> Cc: stable@vger.kernel.org >> Link: https://learn.microsoft.com/en-us/windows-hardware/design/device-experiences/platform-design-for-modern-standby#low-power-core-silicon-cpu-soc-dram [1] >> Fixes: 9d26d3a8f1b0 ("PCI: Put PCIe ports into D3 during suspend") >> Reported-by: Iain Lane <iain@orangesquash.org.uk> >> Closes: https://forums.lenovo.com/t5/Ubuntu/Z13-can-t-resume-from-suspend-with-external-USB-keyboard/m-p/5217121 >> Signed-off-by: Mario Limonciello <mario.limonciello@amd.com> >> --- >> drivers/pci/quirks.c | 61 ++++++++++++++++++++++++++++++++++++++++++++ >> 1 file changed, 61 insertions(+) >> >> diff --git a/drivers/pci/quirks.c b/drivers/pci/quirks.c >> index eeec1d6f9023..ebc0afbc814e 100644 >> --- a/drivers/pci/quirks.c >> +++ b/drivers/pci/quirks.c >> @@ -6188,3 +6188,64 @@ DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x9a31, dpc_log_size); >> DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_XILINX, 0x5020, of_pci_make_dev_node); >> DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_XILINX, 0x5021, of_pci_make_dev_node); >> DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_REDHAT, 0x0005, of_pci_make_dev_node); >> + >> +/* >> + * When AMD PCIe root ports with AMD USB4 controllers attached to them are put >> + * into D3hot or D3cold downstream USB devices may fail to wakeup the system. >> + * This manifests as a missing wakeup interrupt. >> + * >> + * Prevent the associated root port from using PME to wake from D3hot or >> + * D3cold power states during s2idle. >> + * This will effectively put the root port into D0 power state over s2idle. >> + */ >> +static bool child_has_amd_usb4(struct pci_dev *pdev) >> +{ >> + struct pci_dev *child = NULL; >> + >> + while ((child = pci_get_class(PCI_CLASS_SERIAL_USB_USB4, child))) { >> + if (child->vendor != PCI_VENDOR_ID_AMD) >> + continue; >> + if (pcie_find_root_port(child) != pdev) >> + continue; >> + return true; >> + } >> + >> + return false; >> +} >> + >> +static void quirk_reenable_pme(struct pci_dev *dev) >> +{ >> + u16 pmc; >> + >> + if (!dev->pm_cap) >> + return; >> + >> + if (!child_has_amd_usb4(dev)) >> + return; >> + >> + pci_read_config_word(dev, dev->pm_cap + PCI_PM_PMC, &pmc); >> + pmc &= PCI_PM_CAP_PME_MASK; >> + dev->pme_support = pmc >> PCI_PM_CAP_PME_SHIFT; >> +} >> + >> +static void quirk_disable_pme_suspend(struct pci_dev *dev) >> +{ >> + int mask; >> + >> + if (pm_suspend_via_firmware()) >> + return; > > There's always something more to confuse me. Why does > pm_suspend_via_firmware() matter? I can sort of see that Linux > platform power management, which uses the PMC, is not the same > as platform firmware being invoked at the end of a system-wide power > management transition to a sleep state. > > I guess this must have something to do with acpi_suspend_begin() and > acpi_hibernation_begin() (the callers of > pm_set_suspend_via_firmware())? > The "why" has to do with the implementation details of how the platform enters and exits hardware sleep and what happens. It's much different than how ACPI S3 works. Most OEM platforms don't support S3, so if it distracts from the issue and quirk I'm fine to drop this check. >> + if (!child_has_amd_usb4(dev)) >> + return; >> + >> + mask = (PCI_PM_CAP_PME_D3hot|PCI_PM_CAP_PME_D3cold) >> PCI_PM_CAP_PME_SHIFT; >> + if (!(dev->pme_support & mask)) >> + return; >> + dev->pme_support &= ~mask; >> + dev_info_once(&dev->dev, "quirk: disabling PME from D3hot and D3cold at suspend\n"); >> +} >> + >> +DECLARE_PCI_FIXUP_SUSPEND(PCI_VENDOR_ID_AMD, 0x14b9, quirk_disable_pme_suspend); >> +DECLARE_PCI_FIXUP_RESUME(PCI_VENDOR_ID_AMD, 0x14b9, quirk_reenable_pme); >> +DECLARE_PCI_FIXUP_SUSPEND(PCI_VENDOR_ID_AMD, 0x14eb, quirk_disable_pme_suspend); >> +DECLARE_PCI_FIXUP_RESUME(PCI_VENDOR_ID_AMD, 0x14eb, quirk_reenable_pme); >> -- >> 2.34.1 >>
On Sun, Sep 17, 2023 at 04:40:15PM -0500, Bjorn Helgaas wrote: > On Sat, Sep 16, 2023 at 03:36:50PM +0200, Lukas Wunner wrote: > > The "pcie_find_root_port(child) != pdev" check is always false: > > If we were using pci_walk_bus() and only looking at devices below > pdev, I would agree, but since we're using pci_get_class(), which > searches all PCI devices in the system, I'm confused about why it > would always be false. Right, I had misread the patch, I indeed thought it was using pci_walk_bus(). My apologies for the noise. (I guess I assumed that because a global search (as done here) is more expensive than just searching below the Root Port.) Thanks, Lukas
On Thu, 14 Sep 2023, Mario Limonciello wrote: > Iain reports that USB devices can't be used to wake a Lenovo Z13 > from suspend. This problem occurs because the PCIe root port has been put > into D3hot and AMD's platform can't handle USB devices waking the platform > from a hardware sleep state in this case. The platform is put into > a hardware sleep state by the actions of the amd-pmc driver. > > Although the issue is initially reported on a single model it actually > affects all Yellow Carp (Rembrandt) and Pink Sardine (Phoenix) SoCs. > This problem only occurs on Linux specifically when attempting to > wake the platform from a hardware sleep state. > Comparing the behavior on Windows and Linux, Windows doesn't put > the root ports into D3 at this time. > > Linux decides the target state to put the device into at suspend by > this policy: > 1. If platform_pci_power_manageable(): > Use platform_pci_choose_state() > 2. If the device is armed for wakeup: > Select the deepest D-state that supports a PME. > 3. Else: > Use D3hot. > > Devices are considered power manageable by the platform when they have > one or more objects described in the table in section 7.3 of the ACPI 6.5 > specification [1]. In this case the root ports are not power manageable. > > If devices are not considered power manageable; specs are ambiguous as > to what should happen. In this situation Windows 11 puts PCIe ports > in D0 ostensibly due the policy from the "uPEP driver" which is a > complimentary driver the Linux "amd-pmc" driver. > > Linux chooses to allow D3 for these root ports due to the policy > introduced by commit 9d26d3a8f1b0 ("PCI: Put PCIe ports into D3 during > suspend"). > > The Windows uPEP driver expresses the desired state that should be > selected for suspend but Linux doesn't, so introduce a quirk for the > problematic root ports. > > The quirk removes PME support for D3hot and D3cold at suspend time if the > system will be using s2idle. When the port is configured for wakeup this > will prevent these states from being selected in pci_target_state(). > > After the system is resumes the PME support is re-read from the PM > capabilities register to allow opportunistic power savings at runtime by > letting the root port go into D3hot or D3cold. > > Cc: stable@vger.kernel.org > Link: https://learn.microsoft.com/en-us/windows-hardware/design/device-experiences/platform-design-for-modern-standby#low-power-core-silicon-cpu-soc-dram [1] > Fixes: 9d26d3a8f1b0 ("PCI: Put PCIe ports into D3 during suspend") > Reported-by: Iain Lane <iain@orangesquash.org.uk> > Closes: https://forums.lenovo.com/t5/Ubuntu/Z13-can-t-resume-from-suspend-with-external-USB-keyboard/m-p/5217121 > Signed-off-by: Mario Limonciello <mario.limonciello@amd.com> > --- > drivers/pci/quirks.c | 61 ++++++++++++++++++++++++++++++++++++++++++++ > 1 file changed, 61 insertions(+) > > diff --git a/drivers/pci/quirks.c b/drivers/pci/quirks.c > index eeec1d6f9023..ebc0afbc814e 100644 > --- a/drivers/pci/quirks.c > +++ b/drivers/pci/quirks.c > @@ -6188,3 +6188,64 @@ DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x9a31, dpc_log_size); > DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_XILINX, 0x5020, of_pci_make_dev_node); > DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_XILINX, 0x5021, of_pci_make_dev_node); > DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_REDHAT, 0x0005, of_pci_make_dev_node); > + > +/* > + * When AMD PCIe root ports with AMD USB4 controllers attached to them are put > + * into D3hot or D3cold downstream USB devices may fail to wakeup the system. > + * This manifests as a missing wakeup interrupt. > + * > + * Prevent the associated root port from using PME to wake from D3hot or > + * D3cold power states during s2idle. > + * This will effectively put the root port into D0 power state over s2idle. > + */ > +static bool child_has_amd_usb4(struct pci_dev *pdev) > +{ > + struct pci_dev *child = NULL; > + > + while ((child = pci_get_class(PCI_CLASS_SERIAL_USB_USB4, child))) { > + if (child->vendor != PCI_VENDOR_ID_AMD) > + continue; > + if (pcie_find_root_port(child) != pdev) > + continue; > + return true; > + } > + > + return false; > +} > + > +static void quirk_reenable_pme(struct pci_dev *dev) > +{ > + u16 pmc; > + > + if (!dev->pm_cap) > + return; > + > + if (!child_has_amd_usb4(dev)) > + return; > + > + pci_read_config_word(dev, dev->pm_cap + PCI_PM_PMC, &pmc); > + pmc &= PCI_PM_CAP_PME_MASK; > + dev->pme_support = pmc >> PCI_PM_CAP_PME_SHIFT; FIELD_GET(PCI_PM_CAP_PME_MASK, pmc);
diff --git a/drivers/pci/quirks.c b/drivers/pci/quirks.c index eeec1d6f9023..ebc0afbc814e 100644 --- a/drivers/pci/quirks.c +++ b/drivers/pci/quirks.c @@ -6188,3 +6188,64 @@ DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x9a31, dpc_log_size); DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_XILINX, 0x5020, of_pci_make_dev_node); DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_XILINX, 0x5021, of_pci_make_dev_node); DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_REDHAT, 0x0005, of_pci_make_dev_node); + +/* + * When AMD PCIe root ports with AMD USB4 controllers attached to them are put + * into D3hot or D3cold downstream USB devices may fail to wakeup the system. + * This manifests as a missing wakeup interrupt. + * + * Prevent the associated root port from using PME to wake from D3hot or + * D3cold power states during s2idle. + * This will effectively put the root port into D0 power state over s2idle. + */ +static bool child_has_amd_usb4(struct pci_dev *pdev) +{ + struct pci_dev *child = NULL; + + while ((child = pci_get_class(PCI_CLASS_SERIAL_USB_USB4, child))) { + if (child->vendor != PCI_VENDOR_ID_AMD) + continue; + if (pcie_find_root_port(child) != pdev) + continue; + return true; + } + + return false; +} + +static void quirk_reenable_pme(struct pci_dev *dev) +{ + u16 pmc; + + if (!dev->pm_cap) + return; + + if (!child_has_amd_usb4(dev)) + return; + + pci_read_config_word(dev, dev->pm_cap + PCI_PM_PMC, &pmc); + pmc &= PCI_PM_CAP_PME_MASK; + dev->pme_support = pmc >> PCI_PM_CAP_PME_SHIFT; +} + +static void quirk_disable_pme_suspend(struct pci_dev *dev) +{ + int mask; + + if (pm_suspend_via_firmware()) + return; + + if (!child_has_amd_usb4(dev)) + return; + + mask = (PCI_PM_CAP_PME_D3hot|PCI_PM_CAP_PME_D3cold) >> PCI_PM_CAP_PME_SHIFT; + if (!(dev->pme_support & mask)) + return; + dev->pme_support &= ~mask; + dev_info_once(&dev->dev, "quirk: disabling PME from D3hot and D3cold at suspend\n"); +} + +DECLARE_PCI_FIXUP_SUSPEND(PCI_VENDOR_ID_AMD, 0x14b9, quirk_disable_pme_suspend); +DECLARE_PCI_FIXUP_RESUME(PCI_VENDOR_ID_AMD, 0x14b9, quirk_reenable_pme); +DECLARE_PCI_FIXUP_SUSPEND(PCI_VENDOR_ID_AMD, 0x14eb, quirk_disable_pme_suspend); +DECLARE_PCI_FIXUP_RESUME(PCI_VENDOR_ID_AMD, 0x14eb, quirk_reenable_pme);
Iain reports that USB devices can't be used to wake a Lenovo Z13 from suspend. This problem occurs because the PCIe root port has been put into D3hot and AMD's platform can't handle USB devices waking the platform from a hardware sleep state in this case. The platform is put into a hardware sleep state by the actions of the amd-pmc driver. Although the issue is initially reported on a single model it actually affects all Yellow Carp (Rembrandt) and Pink Sardine (Phoenix) SoCs. This problem only occurs on Linux specifically when attempting to wake the platform from a hardware sleep state. Comparing the behavior on Windows and Linux, Windows doesn't put the root ports into D3 at this time. Linux decides the target state to put the device into at suspend by this policy: 1. If platform_pci_power_manageable(): Use platform_pci_choose_state() 2. If the device is armed for wakeup: Select the deepest D-state that supports a PME. 3. Else: Use D3hot. Devices are considered power manageable by the platform when they have one or more objects described in the table in section 7.3 of the ACPI 6.5 specification [1]. In this case the root ports are not power manageable. If devices are not considered power manageable; specs are ambiguous as to what should happen. In this situation Windows 11 puts PCIe ports in D0 ostensibly due the policy from the "uPEP driver" which is a complimentary driver the Linux "amd-pmc" driver. Linux chooses to allow D3 for these root ports due to the policy introduced by commit 9d26d3a8f1b0 ("PCI: Put PCIe ports into D3 during suspend"). The Windows uPEP driver expresses the desired state that should be selected for suspend but Linux doesn't, so introduce a quirk for the problematic root ports. The quirk removes PME support for D3hot and D3cold at suspend time if the system will be using s2idle. When the port is configured for wakeup this will prevent these states from being selected in pci_target_state(). After the system is resumes the PME support is re-read from the PM capabilities register to allow opportunistic power savings at runtime by letting the root port go into D3hot or D3cold. Cc: stable@vger.kernel.org Link: https://learn.microsoft.com/en-us/windows-hardware/design/device-experiences/platform-design-for-modern-standby#low-power-core-silicon-cpu-soc-dram [1] Fixes: 9d26d3a8f1b0 ("PCI: Put PCIe ports into D3 during suspend") Reported-by: Iain Lane <iain@orangesquash.org.uk> Closes: https://forums.lenovo.com/t5/Ubuntu/Z13-can-t-resume-from-suspend-with-external-USB-keyboard/m-p/5217121 Signed-off-by: Mario Limonciello <mario.limonciello@amd.com> --- drivers/pci/quirks.c | 61 ++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 61 insertions(+)