| Message ID | 20240210-topic-8280_pcie-v2-3-1cef4b606883@linaro.org |
|---|---|
| State | New |
| Headers | show |
| Series | Qualcomm PCIe RC shutdown & reinit | expand |
"Properly" is a noise word that suggests "we're doing it right this time" but doesn't hint at what actually makes this better. On Sat, Feb 10, 2024 at 06:10:07PM +0100, Konrad Dybcio wrote: > Currently, we've only been minimizing the power draw while keeping the > RC up at all times. This is suboptimal, as it draws a whole lot of power > and prevents the SoC from power collapsing. Is "power collapse" a technical term specific to this device, or is there some more common term that could be used? I assume the fact that the RC remains powered precludes some lower power state of the entire SoC? > Implement full shutdown and re-initialization to allow for powering off > the controller. > > This is mainly indended for SC8280XP with a broken power rail setup, > which requires a full RC shutdown/reinit in order to reach SoC-wide > power collapse, but sleeping is generally better than not sleeping and > less destructive suspend can be implemented later for platforms that > support it. s/indended/intended/ > config PCIE_QCOM > bool "Qualcomm PCIe controller (host mode)" > depends on OF && (ARCH_QCOM || COMPILE_TEST) > + depends on QCOM_COMMAND_DB || QCOM_COMMAND_DB=n Just out of curiosity since I'm not a Kconfig expert, what does "depends on X || X=n" mean? I guess it's different from "depends on (QCOM_COMMAND_DB || !QCOM_COMMAND_DB)", which I also see used for QCOM_RPMH? Does this reduce compile testing? I see COMPILE_TEST mentioned in a few other QCOM_COMMAND_DB dependencies. > + ret_l23 = readl_poll_timeout(pcie->parf + PARF_PM_STTS, val, > + val & PM_ENTER_L23, 10000, 100000); Are these timeout values rooted in some PCIe or Qcom spec? Would be nice to have a spec citation or other reason for choosing these values. > + reset_control_assert(res->rst); > + usleep_range(2000, 2500); Ditto, some kind of citation would be nice. Bjorn
On 12.02.2024 22:32, Bjorn Helgaas wrote: > "Properly" is a noise word that suggests "we're doing it right this > time" but doesn't hint at what actually makes this better. > > On Sat, Feb 10, 2024 at 06:10:07PM +0100, Konrad Dybcio wrote: >> Currently, we've only been minimizing the power draw while keeping the >> RC up at all times. This is suboptimal, as it draws a whole lot of power >> and prevents the SoC from power collapsing. > > Is "power collapse" a technical term specific to this device, or is > there some more common term that could be used? I assume the fact > that the RC remains powered precludes some lower power state of the > entire SoC? That's spot on, "power collapse" commonly refers to shutting down as many parts of the SoC as possible, in order to achieve miliwatt-order power draw. > >> Implement full shutdown and re-initialization to allow for powering off >> the controller. >> >> This is mainly indended for SC8280XP with a broken power rail setup, >> which requires a full RC shutdown/reinit in order to reach SoC-wide >> power collapse, but sleeping is generally better than not sleeping and >> less destructive suspend can be implemented later for platforms that >> support it. > > s/indended/intended/ > >> config PCIE_QCOM >> bool "Qualcomm PCIe controller (host mode)" >> depends on OF && (ARCH_QCOM || COMPILE_TEST) >> + depends on QCOM_COMMAND_DB || QCOM_COMMAND_DB=n > > Just out of curiosity since I'm not a Kconfig expert, what does > "depends on X || X=n" mean? "not a module" > > I guess it's different from > "depends on (QCOM_COMMAND_DB || !QCOM_COMMAND_DB)", which I also see > used for QCOM_RPMH? Yep > > Does this reduce compile testing? I see COMPILE_TEST mentioned in a > few other QCOM_COMMAND_DB dependencies. I can add "&& COMPILE_TEST", yeah > >> + ret_l23 = readl_poll_timeout(pcie->parf + PARF_PM_STTS, val, >> + val & PM_ENTER_L23, 10000, 100000); > > Are these timeout values rooted in some PCIe or Qcom spec? Would be > nice to have a spec citation or other reason for choosing these > values. > >> + reset_control_assert(res->rst); >> + usleep_range(2000, 2500); > > Ditto, some kind of citation would be nice. Both are magic values coming from Qualcomm BSP, that we suppose we can safely assume (and that's a two-level assumption at this point, I know..) is going to work fine, as it does so on millions of shipped devices. Maybe Mani or Bjorn A can find something interesting in the documentation. Konrad
On Wed, Feb 14, 2024 at 10:33:19PM +0100, Konrad Dybcio wrote: > On 12.02.2024 22:32, Bjorn Helgaas wrote: > > "Properly" is a noise word that suggests "we're doing it right this > > time" but doesn't hint at what actually makes this better. > > > > On Sat, Feb 10, 2024 at 06:10:07PM +0100, Konrad Dybcio wrote: > >> Currently, we've only been minimizing the power draw while keeping the > >> RC up at all times. This is suboptimal, as it draws a whole lot of power > >> and prevents the SoC from power collapsing. > > > > Is "power collapse" a technical term specific to this device, or is > > there some more common term that could be used? I assume the fact > > that the RC remains powered precludes some lower power state of the > > entire SoC? > > That's spot on, "power collapse" commonly refers to shutting down as many > parts of the SoC as possible, in order to achieve miliwatt-order power draw. I'm pretty sure "power collapse" is a Qualcomm:ism so better to use common terminology as Bjorn suggested, and maybe put the Qualcomm wording in parenthesis or similar. Johan
On 15.02.2024 08:13, Johan Hovold wrote: > On Wed, Feb 14, 2024 at 10:33:19PM +0100, Konrad Dybcio wrote: >> On 12.02.2024 22:32, Bjorn Helgaas wrote: >>> "Properly" is a noise word that suggests "we're doing it right this >>> time" but doesn't hint at what actually makes this better. >>> >>> On Sat, Feb 10, 2024 at 06:10:07PM +0100, Konrad Dybcio wrote: >>>> Currently, we've only been minimizing the power draw while keeping the >>>> RC up at all times. This is suboptimal, as it draws a whole lot of power >>>> and prevents the SoC from power collapsing. >>> >>> Is "power collapse" a technical term specific to this device, or is >>> there some more common term that could be used? I assume the fact >>> that the RC remains powered precludes some lower power state of the >>> entire SoC? >> >> That's spot on, "power collapse" commonly refers to shutting down as many >> parts of the SoC as possible, in order to achieve miliwatt-order power draw. > > I'm pretty sure "power collapse" is a Qualcomm:ism so better to use > common terminology as Bjorn suggested, and maybe put the Qualcomm > wording in parenthesis or similar. Ok, I keep hearing it so much that I had previously assumed it's the standard way of describing it.. I'll reword this. Konrad
On 2/10/2024 10:40 PM, Konrad Dybcio wrote: > Currently, we've only been minimizing the power draw while keeping the > RC up at all times. This is suboptimal, as it draws a whole lot of power > and prevents the SoC from power collapsing. > > Implement full shutdown and re-initialization to allow for powering off > the controller. > > This is mainly indended for SC8280XP with a broken power rail setup, > which requires a full RC shutdown/reinit in order to reach SoC-wide > power collapse, but sleeping is generally better than not sleeping and > less destructive suspend can be implemented later for platforms that > support it. > > Co-developed-by: Bjorn Andersson <quic_bjorande@quicinc.com> > Signed-off-by: Konrad Dybcio <konrad.dybcio@linaro.org> > --- > drivers/pci/controller/dwc/Kconfig | 1 + > drivers/pci/controller/dwc/pcie-qcom.c | 159 ++++++++++++++++++++++++++------- > 2 files changed, 126 insertions(+), 34 deletions(-) > > diff --git a/drivers/pci/controller/dwc/Kconfig b/drivers/pci/controller/dwc/Kconfig > index 8afacc90c63b..4ce266951cb6 100644 > --- a/drivers/pci/controller/dwc/Kconfig > +++ b/drivers/pci/controller/dwc/Kconfig > @@ -268,6 +268,7 @@ config PCIE_DW_PLAT_EP > config PCIE_QCOM > bool "Qualcomm PCIe controller (host mode)" > depends on OF && (ARCH_QCOM || COMPILE_TEST) > + depends on QCOM_COMMAND_DB || QCOM_COMMAND_DB=n > depends on PCI_MSI > select PCIE_DW_HOST > select CRC8 > diff --git a/drivers/pci/controller/dwc/pcie-qcom.c b/drivers/pci/controller/dwc/pcie-qcom.c > index 6a469ed213ce..c807833ee4a7 100644 > --- a/drivers/pci/controller/dwc/pcie-qcom.c > +++ b/drivers/pci/controller/dwc/pcie-qcom.c > @@ -30,13 +30,18 @@ > #include <linux/reset.h> > #include <linux/slab.h> > #include <linux/types.h> > +#include <soc/qcom/cmd-db.h> > > #include "../../pci.h" > #include "pcie-designware.h" > > +#include <dt-bindings/interconnect/qcom,icc.h> > +#include <dt-bindings/interconnect/qcom,rpm-icc.h> > + > /* PARF registers */ > #define PARF_SYS_CTRL 0x00 > #define PARF_PM_CTRL 0x20 > +#define PARF_PM_STTS 0x24 > #define PARF_PCS_DEEMPH 0x34 > #define PARF_PCS_SWING 0x38 > #define PARF_PHY_CTRL 0x40 > @@ -80,7 +85,10 @@ > #define L1_CLK_RMV_DIS BIT(1) > > /* PARF_PM_CTRL register fields */ > -#define REQ_NOT_ENTR_L1 BIT(5) > +#define REQ_NOT_ENTR_L1 BIT(5) /* "Prevent L0->L1" */ > + > +/* PARF_PM_STTS register fields */ > +#define PM_ENTER_L23 BIT(5) > > /* PARF_PCS_DEEMPH register fields */ > #define PCS_DEEMPH_TX_DEEMPH_GEN1(x) FIELD_PREP(GENMASK(21, 16), x) > @@ -122,6 +130,7 @@ > > /* ELBI_SYS_CTRL register fields */ > #define ELBI_SYS_CTRL_LT_ENABLE BIT(0) > +#define ELBI_SYS_CTRL_PME_TURNOFF_MSG BIT(4) > > /* AXI_MSTR_RESP_COMP_CTRL0 register fields */ > #define CFG_REMOTE_RD_REQ_BRIDGE_SIZE_2K 0x4 > @@ -243,6 +252,7 @@ struct qcom_pcie { > const struct qcom_pcie_cfg *cfg; > struct dentry *debugfs; > bool suspended; > + bool soc_is_rpmh; > }; > > #define to_qcom_pcie(x) dev_get_drvdata((x)->dev) > @@ -272,6 +282,24 @@ static int qcom_pcie_start_link(struct dw_pcie *pci) > return 0; > } > > +static int qcom_pcie_stop_link(struct dw_pcie *pci) > +{ > + struct qcom_pcie *pcie = to_qcom_pcie(pci); > + u32 ret_l23, val; > + > + writel(ELBI_SYS_CTRL_PME_TURNOFF_MSG, pcie->elbi + ELBI_SYS_CTRL); > + readl(pcie->elbi + ELBI_SYS_CTRL); > + > + ret_l23 = readl_poll_timeout(pcie->parf + PARF_PM_STTS, val, > + val & PM_ENTER_L23, 10000, 100000); > + if (ret_l23) { > + dev_err(pci->dev, "Failed to enter L2/L3\n"); > + return -ETIMEDOUT; > + } > + > + return 0; > +} > + > static void qcom_pcie_clear_hpc(struct dw_pcie *pci) > { > u16 offset = dw_pcie_find_capability(pci, PCI_CAP_ID_EXP); > @@ -987,9 +1015,19 @@ static void qcom_pcie_host_post_init_2_7_0(struct qcom_pcie *pcie) > 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; > + u32 val; > + > + /* Disable PCIe clocks and resets */ > + val = readl(pcie->parf + PARF_PHY_CTRL); > + val |= PHY_TEST_PWR_DOWN; > + writel(val, pcie->parf + PARF_PHY_CTRL); > + readl(pcie->parf + PARF_PHY_CTRL); > > clk_bulk_disable_unprepare(res->num_clks, res->clks); > > + reset_control_assert(res->rst); > + usleep_range(2000, 2500); > + > regulator_bulk_disable(ARRAY_SIZE(res->supplies), res->supplies); > } > > @@ -1545,6 +1583,9 @@ static int qcom_pcie_probe(struct platform_device *pdev) > goto err_phy_exit; > } > > + /* If the soc features RPMh, cmd_db must have been prepared by now */ > + pcie->soc_is_rpmh = !cmd_db_ready(); > + > qcom_pcie_icc_update(pcie); > > if (pcie->mhi) > @@ -1561,58 +1602,108 @@ static int qcom_pcie_probe(struct platform_device *pdev) > return ret; > } > > -static int qcom_pcie_suspend_noirq(struct device *dev) > +static int qcom_pcie_resume_noirq(struct device *dev) > { > struct qcom_pcie *pcie = dev_get_drvdata(dev); > int ret; > > - /* > - * Set minimum bandwidth required to keep data path functional during > - * suspend. > - */ > - ret = icc_set_bw(pcie->icc_mem, 0, kBps_to_icc(1)); > - if (ret) { > - dev_err(dev, "Failed to set interconnect bandwidth: %d\n", ret); > - return ret; > + if (pcie->soc_is_rpmh) { > + /* > + * Undo the tag change from qcom_pcie_suspend_noirq first in > + * case RPMh spontaneously decides to power collapse the > + * platform based on other inputs. > + */ > + icc_set_tag(pcie->icc_mem, QCOM_ICC_TAG_ALWAYS); > + > + /* Flush the tag change */ > + ret = icc_enable(pcie->icc_mem); > + if (ret) { > + dev_err(pcie->pci->dev, "failed to icc_enable: %d", ret); > + return ret; > + } > } > > - /* > - * Turn OFF the resources only for controllers without active PCIe > - * devices. For controllers with active devices, the resources are kept > - * ON and the link is expected to be in L0/L1 (sub)states. > - * > - * Turning OFF the resources for controllers with active PCIe devices > - * will trigger access violation during the end of the suspend cycle, > - * as kernel tries to access the PCIe devices config space for masking > - * MSIs. > - * > - * Also, it is not desirable to put the link into L2/L3 state as that > - * implies VDD supply will be removed and the devices may go into > - * powerdown state. This will affect the lifetime of the storage devices > - * like NVMe. > - */ > - if (!dw_pcie_link_up(pcie->pci)) { > - qcom_pcie_host_deinit(&pcie->pci->pp); > - pcie->suspended = true; > - } > + /* Only check this now to make sure the icc tag has been set. */ > + if (!pcie->suspended) > + return 0; > + > + ret = qcom_pcie_host_init(&pcie->pci->pp); > + if (ret) > + goto revert_icc_tag; > + > + dw_pcie_setup_rc(&pcie->pci->pp); > + > + ret = qcom_pcie_start_link(pcie->pci); > + if (ret) > + goto deinit_host; > + > + /* Ignore the retval, the devices may come up later. */ > + dw_pcie_wait_for_link(pcie->pci); > + > + qcom_pcie_icc_update(pcie); > + > + pcie->suspended = false; > > return 0; > + > +deinit_host: > + qcom_pcie_host_deinit(&pcie->pci->pp); > +revert_icc_tag: > + if (pcie->soc_is_rpmh) { > + icc_set_tag(pcie->icc_mem, QCOM_ICC_TAG_WAKE); > + > + /* Ignore the retval, failing here would be tragic anyway.. */ > + icc_enable(pcie->icc_mem); > + } > + > + return ret; > } > > -static int qcom_pcie_resume_noirq(struct device *dev) > +static int qcom_pcie_suspend_noirq(struct device *dev) > { > struct qcom_pcie *pcie = dev_get_drvdata(dev); > int ret; > > - if (pcie->suspended) { > - ret = qcom_pcie_host_init(&pcie->pci->pp); > + if (pcie->suspended) > + return 0; > + > + if (dw_pcie_link_up(pcie->pci)) { > + ret = qcom_pcie_stop_link(pcie->pci); > if (ret) > return ret; > + } > > - pcie->suspended = false; > + qcom_pcie_host_deinit(&pcie->pci->pp); > + > + if (pcie->soc_is_rpmh) { > + /* > + * The PCIe RC may be covertly accessed by the secure firmware > + * on sleep exit. Use the WAKE bucket to let RPMh pull the plug > + * on PCIe in sleep, but guarantee it comes back up for resume. > + */ > + icc_set_tag(pcie->icc_mem, QCOM_ICC_TAG_WAKE); > + > + /* Flush the tag change */ > + ret = icc_enable(pcie->icc_mem); > + if (ret) { > + dev_err(pcie->pci->dev, "failed to icc_enable %d\n", ret); > + > + /* Revert everything and pray icc calls succeed */ > + return qcom_pcie_resume_noirq(dev); > + } > + } else { > + /* > + * Set minimum bandwidth required to keep data path functional > + * during suspend. > + */ calling qcom_pcie_host_deinit(&pcie->pci->pp) above will turn off all the resources, setting BW to 1Kbps will not make sense here. - Krishna Chaitanya. > + ret = icc_set_bw(pcie->icc_mem, 0, kBps_to_icc(1)); > + if (ret) { > + dev_err(dev, "Failed to set interconnect bandwidth: %d\n", ret); > + return ret; > + } > } > > - qcom_pcie_icc_update(pcie); > + pcie->suspended = true; > > return 0; > } >
On 20.02.2024 5:12 AM, Krishna Chaitanya Chundru wrote: > > > On 2/10/2024 10:40 PM, Konrad Dybcio wrote: >> Currently, we've only been minimizing the power draw while keeping the >> RC up at all times. This is suboptimal, as it draws a whole lot of power >> and prevents the SoC from power collapsing. >> >> Implement full shutdown and re-initialization to allow for powering off >> the controller. >> >> This is mainly indended for SC8280XP with a broken power rail setup, >> which requires a full RC shutdown/reinit in order to reach SoC-wide >> power collapse, but sleeping is generally better than not sleeping and >> less destructive suspend can be implemented later for platforms that >> support it. >> >> Co-developed-by: Bjorn Andersson <quic_bjorande@quicinc.com> >> Signed-off-by: Konrad Dybcio <konrad.dybcio@linaro.org> >> --- [...] >> + if (pcie->soc_is_rpmh) { >> + /* >> + * The PCIe RC may be covertly accessed by the secure firmware >> + * on sleep exit. Use the WAKE bucket to let RPMh pull the plug >> + * on PCIe in sleep, but guarantee it comes back up for resume. >> + */ >> + icc_set_tag(pcie->icc_mem, QCOM_ICC_TAG_WAKE); >> + >> + /* Flush the tag change */ >> + ret = icc_enable(pcie->icc_mem); >> + if (ret) { >> + dev_err(pcie->pci->dev, "failed to icc_enable %d\n", ret); >> + >> + /* Revert everything and pray icc calls succeed */ >> + return qcom_pcie_resume_noirq(dev); >> + } >> + } else { >> + /* >> + * Set minimum bandwidth required to keep data path functional >> + * during suspend. >> + */ > calling qcom_pcie_host_deinit(&pcie->pci->pp) above will turn off all the resources, setting BW to 1Kbps will not make sense here. This is preserving the current behavior, it may be revised later. See ad9b9b6e36c9 ("PCI: qcom: Add support for system suspend and resume") that introduced it, in a perhaps overly 8280-centric fashion. Konrad
diff --git a/drivers/pci/controller/dwc/Kconfig b/drivers/pci/controller/dwc/Kconfig index 8afacc90c63b..4ce266951cb6 100644 --- a/drivers/pci/controller/dwc/Kconfig +++ b/drivers/pci/controller/dwc/Kconfig @@ -268,6 +268,7 @@ config PCIE_DW_PLAT_EP config PCIE_QCOM bool "Qualcomm PCIe controller (host mode)" depends on OF && (ARCH_QCOM || COMPILE_TEST) + depends on QCOM_COMMAND_DB || QCOM_COMMAND_DB=n depends on PCI_MSI select PCIE_DW_HOST select CRC8 diff --git a/drivers/pci/controller/dwc/pcie-qcom.c b/drivers/pci/controller/dwc/pcie-qcom.c index 6a469ed213ce..c807833ee4a7 100644 --- a/drivers/pci/controller/dwc/pcie-qcom.c +++ b/drivers/pci/controller/dwc/pcie-qcom.c @@ -30,13 +30,18 @@ #include <linux/reset.h> #include <linux/slab.h> #include <linux/types.h> +#include <soc/qcom/cmd-db.h> #include "../../pci.h" #include "pcie-designware.h" +#include <dt-bindings/interconnect/qcom,icc.h> +#include <dt-bindings/interconnect/qcom,rpm-icc.h> + /* PARF registers */ #define PARF_SYS_CTRL 0x00 #define PARF_PM_CTRL 0x20 +#define PARF_PM_STTS 0x24 #define PARF_PCS_DEEMPH 0x34 #define PARF_PCS_SWING 0x38 #define PARF_PHY_CTRL 0x40 @@ -80,7 +85,10 @@ #define L1_CLK_RMV_DIS BIT(1) /* PARF_PM_CTRL register fields */ -#define REQ_NOT_ENTR_L1 BIT(5) +#define REQ_NOT_ENTR_L1 BIT(5) /* "Prevent L0->L1" */ + +/* PARF_PM_STTS register fields */ +#define PM_ENTER_L23 BIT(5) /* PARF_PCS_DEEMPH register fields */ #define PCS_DEEMPH_TX_DEEMPH_GEN1(x) FIELD_PREP(GENMASK(21, 16), x) @@ -122,6 +130,7 @@ /* ELBI_SYS_CTRL register fields */ #define ELBI_SYS_CTRL_LT_ENABLE BIT(0) +#define ELBI_SYS_CTRL_PME_TURNOFF_MSG BIT(4) /* AXI_MSTR_RESP_COMP_CTRL0 register fields */ #define CFG_REMOTE_RD_REQ_BRIDGE_SIZE_2K 0x4 @@ -243,6 +252,7 @@ struct qcom_pcie { const struct qcom_pcie_cfg *cfg; struct dentry *debugfs; bool suspended; + bool soc_is_rpmh; }; #define to_qcom_pcie(x) dev_get_drvdata((x)->dev) @@ -272,6 +282,24 @@ static int qcom_pcie_start_link(struct dw_pcie *pci) return 0; } +static int qcom_pcie_stop_link(struct dw_pcie *pci) +{ + struct qcom_pcie *pcie = to_qcom_pcie(pci); + u32 ret_l23, val; + + writel(ELBI_SYS_CTRL_PME_TURNOFF_MSG, pcie->elbi + ELBI_SYS_CTRL); + readl(pcie->elbi + ELBI_SYS_CTRL); + + ret_l23 = readl_poll_timeout(pcie->parf + PARF_PM_STTS, val, + val & PM_ENTER_L23, 10000, 100000); + if (ret_l23) { + dev_err(pci->dev, "Failed to enter L2/L3\n"); + return -ETIMEDOUT; + } + + return 0; +} + static void qcom_pcie_clear_hpc(struct dw_pcie *pci) { u16 offset = dw_pcie_find_capability(pci, PCI_CAP_ID_EXP); @@ -987,9 +1015,19 @@ static void qcom_pcie_host_post_init_2_7_0(struct qcom_pcie *pcie) 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; + u32 val; + + /* Disable PCIe clocks and resets */ + val = readl(pcie->parf + PARF_PHY_CTRL); + val |= PHY_TEST_PWR_DOWN; + writel(val, pcie->parf + PARF_PHY_CTRL); + readl(pcie->parf + PARF_PHY_CTRL); clk_bulk_disable_unprepare(res->num_clks, res->clks); + reset_control_assert(res->rst); + usleep_range(2000, 2500); + regulator_bulk_disable(ARRAY_SIZE(res->supplies), res->supplies); } @@ -1545,6 +1583,9 @@ static int qcom_pcie_probe(struct platform_device *pdev) goto err_phy_exit; } + /* If the soc features RPMh, cmd_db must have been prepared by now */ + pcie->soc_is_rpmh = !cmd_db_ready(); + qcom_pcie_icc_update(pcie); if (pcie->mhi) @@ -1561,58 +1602,108 @@ static int qcom_pcie_probe(struct platform_device *pdev) return ret; } -static int qcom_pcie_suspend_noirq(struct device *dev) +static int qcom_pcie_resume_noirq(struct device *dev) { struct qcom_pcie *pcie = dev_get_drvdata(dev); int ret; - /* - * Set minimum bandwidth required to keep data path functional during - * suspend. - */ - ret = icc_set_bw(pcie->icc_mem, 0, kBps_to_icc(1)); - if (ret) { - dev_err(dev, "Failed to set interconnect bandwidth: %d\n", ret); - return ret; + if (pcie->soc_is_rpmh) { + /* + * Undo the tag change from qcom_pcie_suspend_noirq first in + * case RPMh spontaneously decides to power collapse the + * platform based on other inputs. + */ + icc_set_tag(pcie->icc_mem, QCOM_ICC_TAG_ALWAYS); + + /* Flush the tag change */ + ret = icc_enable(pcie->icc_mem); + if (ret) { + dev_err(pcie->pci->dev, "failed to icc_enable: %d", ret); + return ret; + } } - /* - * Turn OFF the resources only for controllers without active PCIe - * devices. For controllers with active devices, the resources are kept - * ON and the link is expected to be in L0/L1 (sub)states. - * - * Turning OFF the resources for controllers with active PCIe devices - * will trigger access violation during the end of the suspend cycle, - * as kernel tries to access the PCIe devices config space for masking - * MSIs. - * - * Also, it is not desirable to put the link into L2/L3 state as that - * implies VDD supply will be removed and the devices may go into - * powerdown state. This will affect the lifetime of the storage devices - * like NVMe. - */ - if (!dw_pcie_link_up(pcie->pci)) { - qcom_pcie_host_deinit(&pcie->pci->pp); - pcie->suspended = true; - } + /* Only check this now to make sure the icc tag has been set. */ + if (!pcie->suspended) + return 0; + + ret = qcom_pcie_host_init(&pcie->pci->pp); + if (ret) + goto revert_icc_tag; + + dw_pcie_setup_rc(&pcie->pci->pp); + + ret = qcom_pcie_start_link(pcie->pci); + if (ret) + goto deinit_host; + + /* Ignore the retval, the devices may come up later. */ + dw_pcie_wait_for_link(pcie->pci); + + qcom_pcie_icc_update(pcie); + + pcie->suspended = false; return 0; + +deinit_host: + qcom_pcie_host_deinit(&pcie->pci->pp); +revert_icc_tag: + if (pcie->soc_is_rpmh) { + icc_set_tag(pcie->icc_mem, QCOM_ICC_TAG_WAKE); + + /* Ignore the retval, failing here would be tragic anyway.. */ + icc_enable(pcie->icc_mem); + } + + return ret; } -static int qcom_pcie_resume_noirq(struct device *dev) +static int qcom_pcie_suspend_noirq(struct device *dev) { struct qcom_pcie *pcie = dev_get_drvdata(dev); int ret; - if (pcie->suspended) { - ret = qcom_pcie_host_init(&pcie->pci->pp); + if (pcie->suspended) + return 0; + + if (dw_pcie_link_up(pcie->pci)) { + ret = qcom_pcie_stop_link(pcie->pci); if (ret) return ret; + } - pcie->suspended = false; + qcom_pcie_host_deinit(&pcie->pci->pp); + + if (pcie->soc_is_rpmh) { + /* + * The PCIe RC may be covertly accessed by the secure firmware + * on sleep exit. Use the WAKE bucket to let RPMh pull the plug + * on PCIe in sleep, but guarantee it comes back up for resume. + */ + icc_set_tag(pcie->icc_mem, QCOM_ICC_TAG_WAKE); + + /* Flush the tag change */ + ret = icc_enable(pcie->icc_mem); + if (ret) { + dev_err(pcie->pci->dev, "failed to icc_enable %d\n", ret); + + /* Revert everything and pray icc calls succeed */ + return qcom_pcie_resume_noirq(dev); + } + } else { + /* + * Set minimum bandwidth required to keep data path functional + * during suspend. + */ + ret = icc_set_bw(pcie->icc_mem, 0, kBps_to_icc(1)); + if (ret) { + dev_err(dev, "Failed to set interconnect bandwidth: %d\n", ret); + return ret; + } } - qcom_pcie_icc_update(pcie); + pcie->suspended = true; return 0; }
Currently, we've only been minimizing the power draw while keeping the RC up at all times. This is suboptimal, as it draws a whole lot of power and prevents the SoC from power collapsing. Implement full shutdown and re-initialization to allow for powering off the controller. This is mainly indended for SC8280XP with a broken power rail setup, which requires a full RC shutdown/reinit in order to reach SoC-wide power collapse, but sleeping is generally better than not sleeping and less destructive suspend can be implemented later for platforms that support it. Co-developed-by: Bjorn Andersson <quic_bjorande@quicinc.com> Signed-off-by: Konrad Dybcio <konrad.dybcio@linaro.org> --- drivers/pci/controller/dwc/Kconfig | 1 + drivers/pci/controller/dwc/pcie-qcom.c | 159 ++++++++++++++++++++++++++------- 2 files changed, 126 insertions(+), 34 deletions(-)