On 8/13/25 21:29, Bjorn Helgaas wrote:

On Tue, Jun 10, 2025 at 11:07:07AM +0200, Christian Bruel wrote:

quoted

Add driver for the STM32MP25 SoC PCIe Gen1 2.5 GT/s and Gen2 5GT/s
controller based on the DesignWare PCIe core.
...

quoted

+struct stm32_pcie {
+	struct dw_pcie pci;
+	struct regmap *regmap;
+	struct reset_control *rst;
+	struct phy *phy;
+	struct clk *clk;
+	struct gpio_desc *perst_gpio;
+	struct gpio_desc *wake_gpio;
+};
+
+static void stm32_pcie_deassert_perst(struct stm32_pcie *stm32_pcie)
+{
+	/* Delay PERST# de-assertion until the power stabilizes */
+	msleep(PCIE_T_PVPERL_MS);
+
+	gpiod_set_value(stm32_pcie->perst_gpio, 0);
+
+	/* Wait for the REFCLK to stabilize */
+	if (stm32_pcie->perst_gpio)
+		msleep(PCIE_T_RRS_READY_MS);

This obviously relies on gpiod_set_value(stm32_pcie->perst_gpio, 0)
being a no-op when perst_gpio == NULL.  Since we're testing perst_gpio
anyway, I think it might be more readable to avoid relying on
gpiod_set_value() being a no-op:

   msleep(PCIE_T_PVPERL_MS);

   if (stm32_pcie->perst_gpio) {
     gpiod_set_value(stm32_pcie->perst_gpio, 0);
     msleep(PCIE_T_RRS_READY_MS);
   }

(And a similar change in stm32_pcie_assert_perst())

But this seems wrong because PCIE_T_PVPERL_MS is a delay before PERST#
is deasserted, but when perst_gpio == NULL, PERST# has already *been*
deasserted.

So it seems like it would make more sense as:

   if (stm32_pcie->perst_gpio) {
     msleep(PCIE_T_PVPERL_MS);
     gpiod_set_value(stm32_pcie->perst_gpio, 0);
   }

   msleep(PCIE_T_RRS_READY_MS);

I agree, grouping pvperl_ms and the PERST# deassertion together makes 
sense. thank you.

quoted

+}
+
+static void stm32_pcie_assert_perst(struct stm32_pcie *stm32_pcie)
+{
+	gpiod_set_value(stm32_pcie->perst_gpio, 1);
+}
+
+static int stm32_pcie_start_link(struct dw_pcie *pci)
+{
+	struct stm32_pcie *stm32_pcie = to_stm32_pcie(pci);
+
+	return regmap_update_bits(stm32_pcie->regmap, SYSCFG_PCIECR,
+				  STM32MP25_PCIECR_LTSSM_EN,
+				  STM32MP25_PCIECR_LTSSM_EN);
+}
+
+static void stm32_pcie_stop_link(struct dw_pcie *pci)
+{
+	struct stm32_pcie *stm32_pcie = to_stm32_pcie(pci);
+
+	regmap_update_bits(stm32_pcie->regmap, SYSCFG_PCIECR,
+			   STM32MP25_PCIECR_LTSSM_EN, 0);
+}
+
+static int stm32_pcie_suspend_noirq(struct device *dev)
+{
+	struct stm32_pcie *stm32_pcie = dev_get_drvdata(dev);
+	int ret;
+
+	ret = dw_pcie_suspend_noirq(&stm32_pcie->pci);
+	if (ret)
+		return ret;
+
+	stm32_pcie_assert_perst(stm32_pcie);
+
+	clk_disable_unprepare(stm32_pcie->clk);
+
+	if (!device_wakeup_path(dev))
+		phy_exit(stm32_pcie->phy);
+
+	return pinctrl_pm_select_sleep_state(dev);
+}
+
+static int stm32_pcie_resume_noirq(struct device *dev)
+{
+	struct stm32_pcie *stm32_pcie = dev_get_drvdata(dev);
+	int ret;
+
+	/*
+	 * The core clock is gated with CLKREQ# from the COMBOPHY REFCLK,
+	 * thus if no device is present, must force it low with an init pinmux
+	 * to be able to access the DBI registers.

What happens on initial probe if no device is present?  I assume we
access DBI registers in the dw_pcie_host_init() path, and it seems
like we'd have the same issue with DBI not being accessible when no
device is present.

Correct, same issue. The magic happens with the PINCTRL init state that 
is automatically called before probe. As I tried to explain in the 
Documentation in the pinctrl patch:

- if ``init`` and ``default`` are defined in the device tree, the "init"
   state is selected before the driver probe and the "default" state is
   selected after the driver probe.

quoted

+	if (!IS_ERR(dev->pins->init_state))
+		ret = pinctrl_select_state(dev->pins->p, dev->pins->init_state);
+	else
+		ret = pinctrl_pm_select_default_state(dev);
+
+	if (ret) {
+		dev_err(dev, "Failed to activate pinctrl pm state: %d\n", ret);
+		return ret;
+	}
+
+	if (!device_wakeup_path(dev)) {
+		ret = phy_init(stm32_pcie->phy);
+		if (ret) {
+			pinctrl_pm_select_default_state(dev);
+			return ret;
+		}
+	}
+
+	ret = clk_prepare_enable(stm32_pcie->clk);
+	if (ret)
+		goto err_phy_exit;
+
+	stm32_pcie_deassert_perst(stm32_pcie);
+
+	ret = dw_pcie_resume_noirq(&stm32_pcie->pci);
+	if (ret)
+		goto err_disable_clk;
+
+	pinctrl_pm_select_default_state(dev);
+
+	return 0;
+
+err_disable_clk:
+	stm32_pcie_assert_perst(stm32_pcie);
+	clk_disable_unprepare(stm32_pcie->clk);
+
+err_phy_exit:
+	phy_exit(stm32_pcie->phy);
+	pinctrl_pm_select_default_state(dev);
+
+	return ret;
+}
+
+static const struct dev_pm_ops stm32_pcie_pm_ops = {
+	NOIRQ_SYSTEM_SLEEP_PM_OPS(stm32_pcie_suspend_noirq,
+				  stm32_pcie_resume_noirq)
+};
+
+static const struct dw_pcie_host_ops stm32_pcie_host_ops = {
+};
+
+static const struct dw_pcie_ops dw_pcie_ops = {
+	.start_link = stm32_pcie_start_link,
+	.stop_link = stm32_pcie_stop_link
+};
+
+static int stm32_add_pcie_port(struct stm32_pcie *stm32_pcie)
+{
+	struct device *dev = stm32_pcie->pci.dev;
+	unsigned int wake_irq;
+	int ret;
+
+	/* Start to enable resources with PERST# asserted */

I guess if device tree doesn't describe PERST#, we assume PERST# is
actually *deasserted* already at this point (because
stm32_pcie_deassert_perst() does nothing other than the delay)?

yes, this also implies that PV is stable

quoted

+	ret = phy_set_mode(stm32_pcie->phy, PHY_MODE_PCIE);
+	if (ret)
+		return ret;
+
+	ret = phy_init(stm32_pcie->phy);
+	if (ret)
+		return ret;
+
+	ret = regmap_update_bits(stm32_pcie->regmap, SYSCFG_PCIECR,
+				 STM32MP25_PCIECR_TYPE_MASK,
+				 STM32MP25_PCIECR_RC);
+	if (ret)
+		goto err_phy_exit;
+
+	stm32_pcie_deassert_perst(stm32_pcie);
+
+	if (stm32_pcie->wake_gpio) {
+		wake_irq = gpiod_to_irq(stm32_pcie->wake_gpio);
+		ret = dev_pm_set_dedicated_wake_irq(dev, wake_irq);
+		if (ret) {
+			dev_err(dev, "Failed to enable wakeup irq %d\n", ret);
+			goto err_assert_perst;
+		}
+		irq_set_irq_type(wake_irq, IRQ_TYPE_EDGE_FALLING);
+	}
+
+	return 0;
+
+err_assert_perst:
+	stm32_pcie_assert_perst(stm32_pcie);
+
+err_phy_exit:
+	phy_exit(stm32_pcie->phy);
+
+	return ret;
+}
+
+static void stm32_remove_pcie_port(struct stm32_pcie *stm32_pcie)
+{
+	dev_pm_clear_wake_irq(stm32_pcie->pci.dev);
+
+	stm32_pcie_assert_perst(stm32_pcie);
+
+	phy_exit(stm32_pcie->phy);
+}
+
+static int stm32_pcie_parse_port(struct stm32_pcie *stm32_pcie)
+{
+	struct device *dev = stm32_pcie->pci.dev;
+	struct device_node *root_port;
+
+	root_port = of_get_next_available_child(dev->of_node, NULL);
+
+	stm32_pcie->phy = devm_of_phy_get(dev, root_port, NULL);
+	if (IS_ERR(stm32_pcie->phy)) {
+		of_node_put(root_port);
+		return dev_err_probe(dev, PTR_ERR(stm32_pcie->phy),
+				     "Failed to get pcie-phy\n");
+	}
+
+	stm32_pcie->perst_gpio = devm_fwnode_gpiod_get(dev, of_fwnode_handle(root_port),
+						       "reset", GPIOD_OUT_HIGH, NULL);
+	if (IS_ERR(stm32_pcie->perst_gpio)) {
+		if (PTR_ERR(stm32_pcie->perst_gpio) != -ENOENT) {
+			of_node_put(root_port);
+			return dev_err_probe(dev, PTR_ERR(stm32_pcie->perst_gpio),
+					     "Failed to get reset GPIO\n");
+		}
+		stm32_pcie->perst_gpio = NULL;

This looks like perst_gpio is optional in device tree?  Is that really
the case?  It seems hard to ensure we have the PCIE_T_PVPERL_MS and
PCIE_T_RRS_READY_MS delays in the right place if we don't have
perst_gpio.

perst_gpio was initially made optional for pre-silicon platforms with 
hooked PHY functions (or no PHY).
It is assumed that the PV and REFCLK signals are already stable, I 
suppose that without perst_gpio, the system must rely on these 
assumptions. But my experience with other models without perst as gpio 
is limited to this.

quoted

+	}
+
+	stm32_pcie->wake_gpio = devm_fwnode_gpiod_get(dev, of_fwnode_handle(root_port),
+						      "wake", GPIOD_IN, NULL);
+
+	if (IS_ERR(stm32_pcie->wake_gpio)) {
+		if (PTR_ERR(stm32_pcie->wake_gpio) != -ENOENT) {
+			of_node_put(root_port);
+			return dev_err_probe(dev, PTR_ERR(stm32_pcie->wake_gpio),
+					     "Failed to get wake GPIO\n");
+		}
+		stm32_pcie->wake_gpio = NULL;
+	}
+
+	of_node_put(root_port);
+
+	return 0;
+}
+
+static int stm32_pcie_probe(struct platform_device *pdev)
+{
+	struct stm32_pcie *stm32_pcie;
+	struct device *dev = &pdev->dev;
+	int ret;
+
+	stm32_pcie = devm_kzalloc(dev, sizeof(*stm32_pcie), GFP_KERNEL);
+	if (!stm32_pcie)
+		return -ENOMEM;
+
+	stm32_pcie->pci.dev = dev;
+	stm32_pcie->pci.ops = &dw_pcie_ops;
+	stm32_pcie->pci.pp.ops = &stm32_pcie_host_ops;
+
+	stm32_pcie->regmap = syscon_regmap_lookup_by_compatible("st,stm32mp25-syscfg");
+	if (IS_ERR(stm32_pcie->regmap))
+		return dev_err_probe(dev, PTR_ERR(stm32_pcie->regmap),
+				     "No syscfg specified\n");
+
+	stm32_pcie->clk = devm_clk_get(dev, NULL);
+	if (IS_ERR(stm32_pcie->clk))
+		return dev_err_probe(dev, PTR_ERR(stm32_pcie->clk),
+				     "Failed to get PCIe clock source\n");
+
+	stm32_pcie->rst = devm_reset_control_get_exclusive(dev, NULL);
+	if (IS_ERR(stm32_pcie->rst))
+		return dev_err_probe(dev, PTR_ERR(stm32_pcie->rst),
+				     "Failed to get PCIe reset\n");
+
+	ret = stm32_pcie_parse_port(stm32_pcie);
+	if (ret)
+		return ret;
+
+	platform_set_drvdata(pdev, stm32_pcie);
+
+	ret = stm32_add_pcie_port(stm32_pcie);
+	if (ret)
+		return ret;
+
+	reset_control_assert(stm32_pcie->rst);
+	reset_control_deassert(stm32_pcie->rst);
+
+	ret = clk_prepare_enable(stm32_pcie->clk);
+	if (ret) {
+		dev_err(dev, "Core clock enable failed %d\n", ret);
+		goto err_remove_port;
+	}
+
+	ret = pm_runtime_set_active(dev);
+	if (ret < 0) {
+		clk_disable_unprepare(stm32_pcie->clk);
+		stm32_remove_pcie_port(stm32_pcie);
+		return dev_err_probe(dev, ret, "Failed to activate runtime PM\n");
+	}
+
+	pm_runtime_no_callbacks(dev);
+
+	ret = devm_pm_runtime_enable(dev);
+	if (ret < 0) {
+		clk_disable_unprepare(stm32_pcie->clk);
+		stm32_remove_pcie_port(stm32_pcie);
+		return dev_err_probe(dev, ret, "Failed to enable runtime PM\n");
+	}
+
+	ret = dw_pcie_host_init(&stm32_pcie->pci.pp);
+	if (ret)
+		goto err_disable_clk;
+
+	if (stm32_pcie->wake_gpio)
+		device_init_wakeup(dev, true);
+
+	return 0;
+
+err_disable_clk:
+	clk_disable_unprepare(stm32_pcie->clk);
+
+err_remove_port:
+	stm32_remove_pcie_port(stm32_pcie);
+
+	return ret;
+}
+
+static void stm32_pcie_remove(struct platform_device *pdev)
+{
+	struct stm32_pcie *stm32_pcie = platform_get_drvdata(pdev);
+	struct dw_pcie_rp *pp = &stm32_pcie->pci.pp;
+
+	if (stm32_pcie->wake_gpio)
+		device_init_wakeup(&pdev->dev, false);
+
+	dw_pcie_host_deinit(pp);
+
+	clk_disable_unprepare(stm32_pcie->clk);
+
+	stm32_remove_pcie_port(stm32_pcie);
+
+	pm_runtime_put_noidle(&pdev->dev);
+}
+
+static const struct of_device_id stm32_pcie_of_match[] = {
+	{ .compatible = "st,stm32mp25-pcie-rc" },
+	{},
+};
+
+static struct platform_driver stm32_pcie_driver = {
+	.probe = stm32_pcie_probe,
+	.remove = stm32_pcie_remove,
+	.driver = {
+		.name = "stm32-pcie",
+		.of_match_table = stm32_pcie_of_match,
+		.pm = &stm32_pcie_pm_ops,
+		.probe_type = PROBE_PREFER_ASYNCHRONOUS,
+	},
+};
+
+module_platform_driver(stm32_pcie_driver);
+
+MODULE_AUTHOR("Christian Bruel [off-list ref]");
+MODULE_DESCRIPTION("STM32MP25 PCIe Controller driver");
+MODULE_LICENSE("GPL");
+MODULE_DEVICE_TABLE(of, stm32_pcie_of_match);

diff --git a/drivers/pci/controller/dwc/pcie-stm32.h b/drivers/pci/controller/dwc/pcie-stm32.h
new file mode 100644
index 000000000000..387112c4e42c
--- /dev/null
+++ b/drivers/pci/controller/dwc/pcie-stm32.h

@@ -0,0 +1,15 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * ST PCIe driver definitions for STM32-MP25 SoC
+ *
+ * Copyright (C) 2025 STMicroelectronics - All Rights Reserved
+ * Author: Christian Bruel <christian.bruel@foss.st.com>
+ */
+
+#define to_stm32_pcie(x)	dev_get_drvdata((x)->dev)
+
+#define STM32MP25_PCIECR_TYPE_MASK	GENMASK(11, 8)
+#define STM32MP25_PCIECR_LTSSM_EN	BIT(2)
+#define STM32MP25_PCIECR_RC		BIT(10)
+
+#define SYSCFG_PCIECR			0x6000

-- 
2.34.1