PCI: imx6: Add suspend/resume support for i.MX6QDL

From: Stefan Eichenberger <[email protected]>

The suspend/resume support is broken on the i.MX6QDL platform. This
patch resets the link upon resuming to recover functionality. It shares
most of the sequences with other i.MX devices but does not touch the critical registers, which might break PCIe. This patch addresses the
same issue as the following downstream commit:
https://github.com/nxp-imx/linux-imx/commit/4e92355e1f79d225ea842511fcfd42b343b32995
In comparison this patch will also reset the device if possible because
the downstream patch alone would still make the ath10k driver crash.
Without this patch suspend/resume will not work if a PCIe device is connected. The kernel will hang on resume and print an error:
ath10k_pci 0000:01:00.0: Unable to change power state from D3hot to D0, device inaccessible 8<--- cut here ---
Unhandled fault: imprecise external abort (0x1406) at 0x0106f944

Signed-off-by Stefan Eichenberger <[email protected]>

---
Changes in v3:
- Added a new flag to the driver data to indicate that the suspend/resume is broken on the i.MX6QDL platform. (Frank)
- Fix comments to be more relevant (Mani)
- Use imx_pcie_assert_core_reset in suspend (Mani)

 drivers/pci/controller/dwc/pci-imx6.c | 57 +++++++++++++++++++++------
 1 file changed, 46 insertions(+), 11 deletions(-)

drivers/pci/controller/dwc/pci-imx6.c +5 -0

index 808d1f1054173..09e3b15f0908a 100644
@@ -82,6 +82,11 @@ enum imx_pcie_variants {
 #define IMX_PCIE_FLAG_HAS_SERDES		BIT(6)
 #define IMX_PCIE_FLAG_SUPPORT_64BIT		BIT(7)
 #define IMX_PCIE_FLAG_CPU_ADDR_FIXUP		BIT(8)
+/**
+ * Because of ERR005723 (PCIe does not support L2 power down) we need to
+ * workaround suspend resume on some devices which are affected by this errata.
+ */
+#define IMX_PCIE_FLAG_BROKEN_SUSPEND		BIT(9)
 
 #define imx_check_flag(pci, val)	(pci->drvdata->flags & val)
 

drivers/pci/controller/dwc/pci-imx6.c +12 -5

+		 * The minimum for a workaround would be to set PERST# and to
+		 * set the PCIE_TEST_PD flag. However, we can also disable the
+		 * clock which saves some power.
+		 */
+		imx_pcie_assert_core_reset(imx_pcie);
+		imx_pcie->drvdata->enable_ref_clk(imx_pcie, false);
+	} else {
+		imx_pcie_pm_turnoff(imx_pcie);
+		imx_pcie_stop_link(imx_pcie->pci);
+		imx_pcie_host_exit(pp);
+	}
 
 	return 0;
 }

@@ -1253,14 +1268,32 @@ static int imx_pcie_resume_noirq(struct device *dev)
 	if (!(imx_pcie->drvdata->flags & IMX_PCIE_FLAG_SUPPORTS_SUSPEND))
 		return 0;
 
-	ret = imx_pcie_host_init(pp);
-	if (ret)
-		return ret;
-	imx_pcie_msi_save_restore(imx_pcie, false);
-	dw_pcie_setup_rc(pp);
+	if (imx_check_flag(imx_pcie, IMX_PCIE_FLAG_BROKEN_SUSPEND)) {

drivers/pci/controller/dwc/pci-imx6.c +17 -3

+		 */
+		ret = dw_pcie_setup_rc(&imx_pcie->pci->pp);
+		if (ret)
+			return ret;
+		imx_pcie_msi_save_restore(imx_pcie, false);
+	} else {
+		ret = imx_pcie_host_init(pp);
+		if (ret)
+			return ret;
+		imx_pcie_msi_save_restore(imx_pcie, false);
+		dw_pcie_setup_rc(pp);
 
-	if (imx_pcie->link_is_up)
-		imx_pcie_start_link(imx_pcie->pci);
+		if (imx_pcie->link_is_up)
+			imx_pcie_start_link(imx_pcie->pci);
+	}
 
 	return 0;
 }

@@ -1485,7 +1518,9 @@ static const struct imx_pcie_drvdata drvdata[] = {
 	[IMX6Q] = {
 		.variant = IMX6Q,
 		.flags = IMX_PCIE_FLAG_IMX_PHY |
-			 IMX_PCIE_FLAG_IMX_SPEED_CHANGE,
+			 IMX_PCIE_FLAG_IMX_SPEED_CHANGE |
+			 IMX_PCIE_FLAG_BROKEN_SUSPEND |
+			 IMX_PCIE_FLAG_SUPPORTS_SUSPEND,
 		.dbi_length = 0x200,
 		.gpr = "fsl,imx6q-iomuxc-gpr",
 		.clk_names = imx6q_clks,

Signature · 1 line

2.43.0

Quoted 15 lines · Stefan Eichenberger

From: Stefan Eichenberger <[email protected]>

The suspend/resume support is broken on the i.MX6QDL platform. This
patch resets the link upon resuming to recover functionality. It shares
most of the sequences with other i.MX devices but does not touch the critical registers, which might break PCIe. This patch addresses the
same issue as the following downstream commit:
https://github.com/nxp-imx/linux-imx/commit/4e92355e1f79d225ea842511fcfd42b343b32995
In comparison this patch will also reset the device if possible because
the downstream patch alone would still make the ath10k driver crash.
Without this patch suspend/resume will not work if a PCIe device is connected. The kernel will hang on resume and print an error:
ath10k_pci 0000:01:00.0: Unable to change power state from D3hot to D0, device inaccessible 8<--- cut here ---
Unhandled fault: imprecise external abort (0x1406) at 0x0106f944

https://chris.beams.io/posts/git-commit/
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/Documentation/process/maintainer-tip.rst?id=v6.11#n134

Add blank lines between paragraphs. Drop the "8<--- cut here" thing.

What does "reset the link" mean? Please use the same terminology as the PCIe spec when possible.

The downstream commit log ("WARNING: this is not the official workaround; user should take own risk to use it") doesn't exactly inspire confidence.

It sounds like this resets *endpoints*? That sounds scary and unexpected in suspend/resume.

Quoted 31 lines

Signed-off-by Stefan Eichenberger <[email protected]>

---
Changes in v3:
- Added a new flag to the driver data to indicate that the suspend/resume is broken on the i.MX6QDL platform. (Frank)
- Fix comments to be more relevant (Mani)
- Use imx_pcie_assert_core_reset in suspend (Mani)

 drivers/pci/controller/dwc/pci-imx6.c | 57 +++++++++++++++++++++------
 1 file changed, 46 insertions(+), 11 deletions(-)

drivers/pci/controller/dwc/pci-imx6.c +5 -0

index 808d1f1054173..09e3b15f0908a 100644
@@ -82,6 +82,11 @@ enum imx_pcie_variants {
 #define IMX_PCIE_FLAG_HAS_SERDES		BIT(6)
 #define IMX_PCIE_FLAG_SUPPORT_64BIT		BIT(7)
 #define IMX_PCIE_FLAG_CPU_ADDR_FIXUP		BIT(8)
+/**
+ * Because of ERR005723 (PCIe does not support L2 power down) we need to
+ * workaround suspend resume on some devices which are affected by this errata.
+ */
+#define IMX_PCIE_FLAG_BROKEN_SUSPEND		BIT(9)
 
 #define imx_check_flag(pci, val)	(pci->drvdata->flags & val)
 

@@ -1237,9 +1242,19 @@ static int imx_pcie_suspend_noirq(struct device *dev)
 		return 0;
 

drivers/pci/controller/dwc/pci-imx6.c static int imx_pcie_suspend_noirq(struct device *dev) · line 1246

Quoted 6 lines

drivers/pci/controller/dwc/pci-imx6.c +2 -3

 	imx_pcie_msi_save_restore(imx_pcie, true);
-	imx_pcie_pm_turnoff(imx_pcie);
-	imx_pcie_stop_link(imx_pcie->pci);
-	imx_pcie_host_exit(pp);
+	if (imx_check_flag(imx_pcie, IMX_PCIE_FLAG_BROKEN_SUSPEND)) {
+		/**

Single asterisks above, here, and below.

Quoted 28 lines

drivers/pci/controller/dwc/pci-imx6.c +16 -5

+		 * The minimum for a workaround would be to set PERST# and to
+		 * set the PCIE_TEST_PD flag. However, we can also disable the
+		 * clock which saves some power.
+		 */
+		imx_pcie_assert_core_reset(imx_pcie);
+		imx_pcie->drvdata->enable_ref_clk(imx_pcie, false);
+	} else {
+		imx_pcie_pm_turnoff(imx_pcie);
+		imx_pcie_stop_link(imx_pcie->pci);
+		imx_pcie_host_exit(pp);
+	}
 
 	return 0;
 }

@@ -1253,14 +1268,32 @@ static int imx_pcie_resume_noirq(struct device *dev)
 	if (!(imx_pcie->drvdata->flags & IMX_PCIE_FLAG_SUPPORTS_SUSPEND))
 		return 0;
 
-	ret = imx_pcie_host_init(pp);
-	if (ret)
-		return ret;
-	imx_pcie_msi_save_restore(imx_pcie, false);
-	dw_pcie_setup_rc(pp);
+	if (imx_check_flag(imx_pcie, IMX_PCIE_FLAG_BROKEN_SUSPEND)) {
+		ret = imx_pcie->drvdata->enable_ref_clk(imx_pcie, true);
+		if (ret)
+			return ret;
+		ret = imx_pcie_deassert_core_reset(imx_pcie);

drivers/pci/controller/dwc/pci-imx6.c static int imx_pcie_resume_noirq(struct device *dev) · line 1281

Quoted 6 lines

drivers/pci/controller/dwc/pci-imx6.c +6 -0

+		if (ret)
+			return ret;
+		/**
+		 * Using PCIE_TEST_PD seems to disable msi and powers down the
+		 * root complex. This is why we have to setup the rc again and
+		 * why we have to restore the msi register.

s/msi/MSI/

Quoted 19 lines

drivers/pci/controller/dwc/pci-imx6.c +14 -2

+		 */
+		ret = dw_pcie_setup_rc(&imx_pcie->pci->pp);
+		if (ret)
+			return ret;
+		imx_pcie_msi_save_restore(imx_pcie, false);
+	} else {
+		ret = imx_pcie_host_init(pp);
+		if (ret)
+			return ret;
+		imx_pcie_msi_save_restore(imx_pcie, false);
+		dw_pcie_setup_rc(pp);
 
-	if (imx_pcie->link_is_up)
-		imx_pcie_start_link(imx_pcie->pci);
+		if (imx_pcie->link_is_up)
+			imx_pcie_start_link(imx_pcie->pci);
+	}
 

drivers/pci/controller/dwc/pci-imx6.c static const struct imx_pcie_drvdata drvdata[] = { · line 1526

Quoted 13 lines

drivers/pci/controller/dwc/pci-imx6.c +3 -1

 	return 0;
 }

@@ -1485,7 +1518,9 @@ static const struct imx_pcie_drvdata drvdata[] = {
 	[IMX6Q] = {
 		.variant = IMX6Q,
 		.flags = IMX_PCIE_FLAG_IMX_PHY |
-			 IMX_PCIE_FLAG_IMX_SPEED_CHANGE,
+			 IMX_PCIE_FLAG_IMX_SPEED_CHANGE |
+			 IMX_PCIE_FLAG_BROKEN_SUSPEND |
+			 IMX_PCIE_FLAG_SUPPORTS_SUSPEND,
 		.dbi_length = 0x200,
 		.gpr = "fsl,imx6q-iomuxc-gpr",
 		.clk_names = imx6q_clks,

Quoted 14 lines · Bjorn Helgaas

Quoted 13 lines · Stefan Eichenberger

From: Stefan Eichenberger <[email protected]>

The suspend/resume support is broken on the i.MX6QDL platform. This
patch resets the link upon resuming to recover functionality. It shares
most of the sequences with other i.MX devices but does not touch the critical registers, which might break PCIe. This patch addresses the
same issue as the following downstream commit:
https://github.com/nxp-imx/linux-imx/commit/4e92355e1f79d225ea842511fcfd42b343b32995
In comparison this patch will also reset the device if possible because
the downstream patch alone would still make the ath10k driver crash.
Without this patch suspend/resume will not work if a PCIe device is connected. The kernel will hang on resume and print an error:
ath10k_pci 0000:01:00.0: Unable to change power state from D3hot to D0, device inaccessible

...

Quoted 6 lines

The downstream commit log ("WARNING: this is not the official workaround; user should take own risk to use it") doesn't exactly inspire confidence.

It sounds like this resets *endpoints*? That sounds scary and unexpected in suspend/resume.

Yes, I completely agree with you, but NXP has never come up with an "official" workaround. Our problem is that with the current implementation, suspend/resume is completely broken when a PCIe device is connected. With this proposed patch we at least have a working device after resume. Even for the other i.MX devices, the driver resets the endpoints in the resume function (imx_pcie_resume_noir -> imx_pcie_host_init -> imx_pcie_assert_core_reset), we just do that now for the i.MX6QDL as well. If it is more appropriate to call imx_pcie_assert_core_reset in resume as we do for the other devices, that would be fine with me as well. I was thinking that if we need to reset the device anyway, we could put it into reset on suspend, as this might save some extra power.