| Message ID | 20231010155914.9516-3-manivannan.sadhasivam@linaro.org (mailing list archive) |
|---|---|
| State | Accepted |
| Delegated to: | Krzysztof Wilczyński |
| Headers | show |
| Series | PCI: qcom: Enable ASPM on host bridge and devices | expand |
On Tue, Oct 10, 2023 at 09:29:14PM +0530, Manivannan Sadhasivam wrote: > ASPM is supported by Qcom host controllers/bridges on most of the recent > platforms and so the devices tested so far. But for enabling ASPM by > default (without Kconfig/cmdline/sysfs), BIOS has to enable ASPM on both > host bridge and downstream devices during boot. Unfortunately, none of the > BIOS available on Qcom platforms enables ASPM. I think this covers over a PCI core defect. If the devices advertise ASPM support, which both the qcom host controller and the endpoint devices do, the PCI core should be able to enable it without being prodded as this patch does. We had a long conversation about this at [1], but never came to a good resolution. Since we don't know how to fix the PCI core issue, I guess we have no choice but to do things like this patch, at least for now. If/when we ever *do* fix the PCI core issue, it would likely result in enabling ASPM (if advertised by both ends of the link) for *all* qcom controllers, not just the 1.9.0 ones. And I think that even today, users can enable ASPM on non-1.9.0 controllers via sysfs. So if you are concerned about ASPM not being tested on those controllers, you may want to make them not advertise ASPM support. Even with this patch, I guess hot-added devices don't get ASPM enabled? That's basically what [1] is about. Bjorn [1] https://lore.kernel.org/linux-pci/20230615070421.1704133-1-kai.heng.feng@canonical.com/ > Due to this, the platforms > making use of Qcom SoCs draw high power during runtime. > > To fix this power issue, users/distros have to enable ASPM using configs > such as (Kconfig/cmdline/sysfs) or the BIOS has to start enabling ASPM. > The latter may happen in the future, but that won't address the issue on > current platforms. Also, asking users/distros to enable a feature to get > the power management right would provide an unpleasant out-of-the-box > experience. > > So the apt solution is to enable ASPM in the controller driver itself. And > this is being accomplished by calling pci_enable_link_state() in the newly > introduced host_post_init() callback for all the devices connected to the > bus. This function enables all supported link low power states for both > host bridge and the downstream devices. > > Due to limited testing, ASPM is only enabled for platforms making use of > ops_1_9_0 callbacks. > > Signed-off-by: Manivannan Sadhasivam <manivannan.sadhasivam@linaro.org> > --- > drivers/pci/controller/dwc/pcie-qcom.c | 28 ++++++++++++++++++++++++++ > 1 file changed, 28 insertions(+) > > diff --git a/drivers/pci/controller/dwc/pcie-qcom.c b/drivers/pci/controller/dwc/pcie-qcom.c > index 367acb419a2b..c324c3daaa5a 100644 > --- a/drivers/pci/controller/dwc/pcie-qcom.c > +++ b/drivers/pci/controller/dwc/pcie-qcom.c > @@ -222,6 +222,7 @@ struct qcom_pcie_ops { > int (*get_resources)(struct qcom_pcie *pcie); > int (*init)(struct qcom_pcie *pcie); > int (*post_init)(struct qcom_pcie *pcie); > + void (*host_post_init)(struct qcom_pcie *pcie); > void (*deinit)(struct qcom_pcie *pcie); > void (*ltssm_enable)(struct qcom_pcie *pcie); > int (*config_sid)(struct qcom_pcie *pcie); > @@ -967,6 +968,22 @@ static int qcom_pcie_post_init_2_7_0(struct qcom_pcie *pcie) > return 0; } > > +static int qcom_pcie_enable_aspm(struct pci_dev *pdev, void *userdata) > +{ > + /* Downstream devices need to be in D0 state before enabling PCI PM substates */ > + pci_set_power_state(pdev, PCI_D0); > + pci_enable_link_state(pdev, PCIE_LINK_STATE_ALL); > + > + return 0; > +} > + > +static void qcom_pcie_host_post_init_2_7_0(struct qcom_pcie *pcie) > +{ > + struct dw_pcie_rp *pp = &pcie->pci->pp; > + > + pci_walk_bus(pp->bridge->bus, qcom_pcie_enable_aspm, NULL); > +} > + > static void qcom_pcie_deinit_2_7_0(struct qcom_pcie *pcie) > { > struct qcom_pcie_resources_2_7_0 *res = &pcie->res.v2_7_0; > @@ -1219,9 +1236,19 @@ static void qcom_pcie_host_deinit(struct dw_pcie_rp *pp) > pcie->cfg->ops->deinit(pcie); > } > > +static void qcom_pcie_host_post_init(struct dw_pcie_rp *pp) > +{ > + struct dw_pcie *pci = to_dw_pcie_from_pp(pp); > + struct qcom_pcie *pcie = to_qcom_pcie(pci); > + > + if (pcie->cfg->ops->host_post_init) > + pcie->cfg->ops->host_post_init(pcie); > +} > + > static const struct dw_pcie_host_ops qcom_pcie_dw_ops = { > .host_init = qcom_pcie_host_init, > .host_deinit = qcom_pcie_host_deinit, > + .host_post_init = qcom_pcie_host_post_init, > }; > > /* Qcom IP rev.: 2.1.0 Synopsys IP rev.: 4.01a */ > @@ -1283,6 +1310,7 @@ static const struct qcom_pcie_ops ops_1_9_0 = { > .get_resources = qcom_pcie_get_resources_2_7_0, > .init = qcom_pcie_init_2_7_0, > .post_init = qcom_pcie_post_init_2_7_0, > + .host_post_init = qcom_pcie_host_post_init_2_7_0, > .deinit = qcom_pcie_deinit_2_7_0, > .ltssm_enable = qcom_pcie_2_3_2_ltssm_enable, > .config_sid = qcom_pcie_config_sid_1_9_0, > -- > 2.25.1 >
On 10/10/23 17:59, Manivannan Sadhasivam wrote: > ASPM is supported by Qcom host controllers/bridges on most of the recent > platforms and so the devices tested so far. But for enabling ASPM by > default (without Kconfig/cmdline/sysfs), BIOS has to enable ASPM on both > host bridge and downstream devices during boot. Unfortunately, none of the > BIOS available on Qcom platforms enables ASPM. Due to this, the platforms > making use of Qcom SoCs draw high power during runtime. > > To fix this power issue, users/distros have to enable ASPM using configs > such as (Kconfig/cmdline/sysfs) or the BIOS has to start enabling ASPM. > The latter may happen in the future, but that won't address the issue on > current platforms. Also, asking users/distros to enable a feature to get > the power management right would provide an unpleasant out-of-the-box > experience. > > So the apt solution is to enable ASPM in the controller driver itself. And > this is being accomplished by calling pci_enable_link_state() in the newly > introduced host_post_init() callback for all the devices connected to the > bus. This function enables all supported link low power states for both > host bridge and the downstream devices. > > Due to limited testing, ASPM is only enabled for platforms making use of > ops_1_9_0 callbacks. > > Signed-off-by: Manivannan Sadhasivam <manivannan.sadhasivam@linaro.org> > --- [...] > +static int qcom_pcie_enable_aspm(struct pci_dev *pdev, void *userdata) > +{ > + /* Downstream devices need to be in D0 state before enabling PCI PM substates */ > + pci_set_power_state(pdev, PCI_D0); > + pci_enable_link_state(pdev, PCIE_LINK_STATE_ALL); Do we not care about retval here? > + > + return 0; > +} > + > +static void qcom_pcie_host_post_init_2_7_0(struct qcom_pcie *pcie) post_init_enable_aspm? Konrad
On Tue, Oct 10, 2023 at 11:29:45AM -0500, Bjorn Helgaas wrote: > On Tue, Oct 10, 2023 at 09:29:14PM +0530, Manivannan Sadhasivam wrote: > > ASPM is supported by Qcom host controllers/bridges on most of the recent > > platforms and so the devices tested so far. But for enabling ASPM by > > default (without Kconfig/cmdline/sysfs), BIOS has to enable ASPM on both > > host bridge and downstream devices during boot. Unfortunately, none of the > > BIOS available on Qcom platforms enables ASPM. > > I think this covers over a PCI core defect. If the devices advertise > ASPM support, which both the qcom host controller and the endpoint > devices do, the PCI core should be able to enable it without being > prodded as this patch does. > Right. > We had a long conversation about this at [1], but never came to a good > resolution. Since we don't know how to fix the PCI core issue, I > guess we have no choice but to do things like this patch, at least for > now. > > If/when we ever *do* fix the PCI core issue, it would likely result in > enabling ASPM (if advertised by both ends of the link) for *all* qcom > controllers, not just the 1.9.0 ones. > That would be a welcome move IMO :) > And I think that even today, users can enable ASPM on non-1.9.0 > controllers via sysfs. So if you are concerned about ASPM not being > tested on those controllers, you may want to make them not advertise > ASPM support. > That will completely take away the power saving if users want it at some time. I'm planning to enable ASPM for other configs as well in the coming days once I get hold of the test platforms. For platforms where ASPM is not really desired, like IPQ SoCs used in routers, I will just disable it. > Even with this patch, I guess hot-added devices don't get ASPM > enabled? That's basically what [1] is about. > Yeah, that's a limitation like the BIOS enablement. Fortunately, most of the platforms based on Qcom SoCs do not have hot pluggable PCIe slots like in PCs. But we should fix it too. - Mani > Bjorn > > [1] https://lore.kernel.org/linux-pci/20230615070421.1704133-1-kai.heng.feng@canonical.com/ > > > Due to this, the platforms > > making use of Qcom SoCs draw high power during runtime. > > > > To fix this power issue, users/distros have to enable ASPM using configs > > such as (Kconfig/cmdline/sysfs) or the BIOS has to start enabling ASPM. > > The latter may happen in the future, but that won't address the issue on > > current platforms. Also, asking users/distros to enable a feature to get > > the power management right would provide an unpleasant out-of-the-box > > experience. > > > > So the apt solution is to enable ASPM in the controller driver itself. And > > this is being accomplished by calling pci_enable_link_state() in the newly > > introduced host_post_init() callback for all the devices connected to the > > bus. This function enables all supported link low power states for both > > host bridge and the downstream devices. > > > > Due to limited testing, ASPM is only enabled for platforms making use of > > ops_1_9_0 callbacks. > > > > Signed-off-by: Manivannan Sadhasivam <manivannan.sadhasivam@linaro.org> > > --- > > drivers/pci/controller/dwc/pcie-qcom.c | 28 ++++++++++++++++++++++++++ > > 1 file changed, 28 insertions(+) > > > > diff --git a/drivers/pci/controller/dwc/pcie-qcom.c b/drivers/pci/controller/dwc/pcie-qcom.c > > index 367acb419a2b..c324c3daaa5a 100644 > > --- a/drivers/pci/controller/dwc/pcie-qcom.c > > +++ b/drivers/pci/controller/dwc/pcie-qcom.c > > @@ -222,6 +222,7 @@ struct qcom_pcie_ops { > > int (*get_resources)(struct qcom_pcie *pcie); > > int (*init)(struct qcom_pcie *pcie); > > int (*post_init)(struct qcom_pcie *pcie); > > + void (*host_post_init)(struct qcom_pcie *pcie); > > void (*deinit)(struct qcom_pcie *pcie); > > void (*ltssm_enable)(struct qcom_pcie *pcie); > > int (*config_sid)(struct qcom_pcie *pcie); > > @@ -967,6 +968,22 @@ static int qcom_pcie_post_init_2_7_0(struct qcom_pcie *pcie) > > return 0; > } > > > > +static int qcom_pcie_enable_aspm(struct pci_dev *pdev, void *userdata) > > +{ > > + /* Downstream devices need to be in D0 state before enabling PCI PM substates */ > > + pci_set_power_state(pdev, PCI_D0); > > + pci_enable_link_state(pdev, PCIE_LINK_STATE_ALL); > > + > > + return 0; > > +} > > + > > +static void qcom_pcie_host_post_init_2_7_0(struct qcom_pcie *pcie) > > +{ > > + struct dw_pcie_rp *pp = &pcie->pci->pp; > > + > > + pci_walk_bus(pp->bridge->bus, qcom_pcie_enable_aspm, NULL); > > +} > > + > > static void qcom_pcie_deinit_2_7_0(struct qcom_pcie *pcie) > > { > > struct qcom_pcie_resources_2_7_0 *res = &pcie->res.v2_7_0; > > @@ -1219,9 +1236,19 @@ static void qcom_pcie_host_deinit(struct dw_pcie_rp *pp) > > pcie->cfg->ops->deinit(pcie); > > } > > > > +static void qcom_pcie_host_post_init(struct dw_pcie_rp *pp) > > +{ > > + struct dw_pcie *pci = to_dw_pcie_from_pp(pp); > > + struct qcom_pcie *pcie = to_qcom_pcie(pci); > > + > > + if (pcie->cfg->ops->host_post_init) > > + pcie->cfg->ops->host_post_init(pcie); > > +} > > + > > static const struct dw_pcie_host_ops qcom_pcie_dw_ops = { > > .host_init = qcom_pcie_host_init, > > .host_deinit = qcom_pcie_host_deinit, > > + .host_post_init = qcom_pcie_host_post_init, > > }; > > > > /* Qcom IP rev.: 2.1.0 Synopsys IP rev.: 4.01a */ > > @@ -1283,6 +1310,7 @@ static const struct qcom_pcie_ops ops_1_9_0 = { > > .get_resources = qcom_pcie_get_resources_2_7_0, > > .init = qcom_pcie_init_2_7_0, > > .post_init = qcom_pcie_post_init_2_7_0, > > + .host_post_init = qcom_pcie_host_post_init_2_7_0, > > .deinit = qcom_pcie_deinit_2_7_0, > > .ltssm_enable = qcom_pcie_2_3_2_ltssm_enable, > > .config_sid = qcom_pcie_config_sid_1_9_0, > > -- > > 2.25.1 > >
On Tue, Oct 10, 2023 at 06:33:52PM +0200, Konrad Dybcio wrote: > > > On 10/10/23 17:59, Manivannan Sadhasivam wrote: > > ASPM is supported by Qcom host controllers/bridges on most of the recent > > platforms and so the devices tested so far. But for enabling ASPM by > > default (without Kconfig/cmdline/sysfs), BIOS has to enable ASPM on both > > host bridge and downstream devices during boot. Unfortunately, none of the > > BIOS available on Qcom platforms enables ASPM. Due to this, the platforms > > making use of Qcom SoCs draw high power during runtime. > > > > To fix this power issue, users/distros have to enable ASPM using configs > > such as (Kconfig/cmdline/sysfs) or the BIOS has to start enabling ASPM. > > The latter may happen in the future, but that won't address the issue on > > current platforms. Also, asking users/distros to enable a feature to get > > the power management right would provide an unpleasant out-of-the-box > > experience. > > > > So the apt solution is to enable ASPM in the controller driver itself. And > > this is being accomplished by calling pci_enable_link_state() in the newly > > introduced host_post_init() callback for all the devices connected to the > > bus. This function enables all supported link low power states for both > > host bridge and the downstream devices. > > > > Due to limited testing, ASPM is only enabled for platforms making use of > > ops_1_9_0 callbacks. > > > > Signed-off-by: Manivannan Sadhasivam <manivannan.sadhasivam@linaro.org> > > --- > [...] > > > +static int qcom_pcie_enable_aspm(struct pci_dev *pdev, void *userdata) > > +{ > > + /* Downstream devices need to be in D0 state before enabling PCI PM substates */ > > + pci_set_power_state(pdev, PCI_D0); > > + pci_enable_link_state(pdev, PCIE_LINK_STATE_ALL); > Do we not care about retval here? > No. Even if it fails, we shouldn't care about it. > > + > > + return 0; > > +} > > + > > +static void qcom_pcie_host_post_init_2_7_0(struct qcom_pcie *pcie) > post_init_enable_aspm? > The scope of this callback may get extended in the future. So I'd keep it as it is. - Mani
diff --git a/drivers/pci/controller/dwc/pcie-qcom.c b/drivers/pci/controller/dwc/pcie-qcom.c index 367acb419a2b..c324c3daaa5a 100644 --- a/drivers/pci/controller/dwc/pcie-qcom.c +++ b/drivers/pci/controller/dwc/pcie-qcom.c @@ -222,6 +222,7 @@ struct qcom_pcie_ops { int (*get_resources)(struct qcom_pcie *pcie); int (*init)(struct qcom_pcie *pcie); int (*post_init)(struct qcom_pcie *pcie); + void (*host_post_init)(struct qcom_pcie *pcie); void (*deinit)(struct qcom_pcie *pcie); void (*ltssm_enable)(struct qcom_pcie *pcie); int (*config_sid)(struct qcom_pcie *pcie); @@ -967,6 +968,22 @@ static int qcom_pcie_post_init_2_7_0(struct qcom_pcie *pcie) return 0; } +static int qcom_pcie_enable_aspm(struct pci_dev *pdev, void *userdata) +{ + /* Downstream devices need to be in D0 state before enabling PCI PM substates */ + pci_set_power_state(pdev, PCI_D0); + pci_enable_link_state(pdev, PCIE_LINK_STATE_ALL); + + return 0; +} + +static void qcom_pcie_host_post_init_2_7_0(struct qcom_pcie *pcie) +{ + struct dw_pcie_rp *pp = &pcie->pci->pp; + + pci_walk_bus(pp->bridge->bus, qcom_pcie_enable_aspm, NULL); +} + static void qcom_pcie_deinit_2_7_0(struct qcom_pcie *pcie) { struct qcom_pcie_resources_2_7_0 *res = &pcie->res.v2_7_0; @@ -1219,9 +1236,19 @@ static void qcom_pcie_host_deinit(struct dw_pcie_rp *pp) pcie->cfg->ops->deinit(pcie); } +static void qcom_pcie_host_post_init(struct dw_pcie_rp *pp) +{ + struct dw_pcie *pci = to_dw_pcie_from_pp(pp); + struct qcom_pcie *pcie = to_qcom_pcie(pci); + + if (pcie->cfg->ops->host_post_init) + pcie->cfg->ops->host_post_init(pcie); +} + static const struct dw_pcie_host_ops qcom_pcie_dw_ops = { .host_init = qcom_pcie_host_init, .host_deinit = qcom_pcie_host_deinit, + .host_post_init = qcom_pcie_host_post_init, }; /* Qcom IP rev.: 2.1.0 Synopsys IP rev.: 4.01a */ @@ -1283,6 +1310,7 @@ static const struct qcom_pcie_ops ops_1_9_0 = { .get_resources = qcom_pcie_get_resources_2_7_0, .init = qcom_pcie_init_2_7_0, .post_init = qcom_pcie_post_init_2_7_0, + .host_post_init = qcom_pcie_host_post_init_2_7_0, .deinit = qcom_pcie_deinit_2_7_0, .ltssm_enable = qcom_pcie_2_3_2_ltssm_enable, .config_sid = qcom_pcie_config_sid_1_9_0,
ASPM is supported by Qcom host controllers/bridges on most of the recent platforms and so the devices tested so far. But for enabling ASPM by default (without Kconfig/cmdline/sysfs), BIOS has to enable ASPM on both host bridge and downstream devices during boot. Unfortunately, none of the BIOS available on Qcom platforms enables ASPM. Due to this, the platforms making use of Qcom SoCs draw high power during runtime. To fix this power issue, users/distros have to enable ASPM using configs such as (Kconfig/cmdline/sysfs) or the BIOS has to start enabling ASPM. The latter may happen in the future, but that won't address the issue on current platforms. Also, asking users/distros to enable a feature to get the power management right would provide an unpleasant out-of-the-box experience. So the apt solution is to enable ASPM in the controller driver itself. And this is being accomplished by calling pci_enable_link_state() in the newly introduced host_post_init() callback for all the devices connected to the bus. This function enables all supported link low power states for both host bridge and the downstream devices. Due to limited testing, ASPM is only enabled for platforms making use of ops_1_9_0 callbacks. Signed-off-by: Manivannan Sadhasivam <manivannan.sadhasivam@linaro.org> --- drivers/pci/controller/dwc/pcie-qcom.c | 28 ++++++++++++++++++++++++++ 1 file changed, 28 insertions(+)