Skip to contentSkip to conversation

dra7xx: PCIe IRQ handling rework

PATCH In review Vignesh R 3 participants 12 messages Feb 9, 2018 – Mar 15, 2018 PCI

lore
Share & tools
lore

Vignesh R

author
#0 lore raw

cover dra7xx: PCIe IRQ handling rework

This series re-works dra7xx MSI IRQ handler to avoid issues of missing of MSI IRQs and Legacy IRQs.

Tested on AM572 EVM with TUSB PCIe USB card and Intel PCIe WiFi cards.

Vignesh R (3):
Revert "PCI: dwc: Clear MSI interrupt status after it is handled, not
before"
PCI: dwc: pci-dra7xx: Improve MSI IRQ handling
PCI: dwc: pci-dra7xx: Handle legacy and MSI IRQs together

 drivers/pci/dwc/pci-dra7xx.c           | 32 +++++++++++++++++++++-----------
 drivers/pci/dwc/pcie-designware-host.c |  2 +-
 2 files changed, 22 insertions(+), 12 deletions(-)
Signature · 1 line
2.16.1
author
#1 lore raw

PATCH 1/3 Revert "PCI: dwc: Clear MSI interrupt status after it is handled, not before"

Since commit 06e15e6883bed ("PCI: dwc: Clear MSI interrupt status after it is handled, not before"), MSI IRQ status in PCIE_MSI_INTR0_STATUS register is cleared after calling EP's IRQ handler.
But, MSI IRQs in case of PCIe are bit like edge interrupts. If a another MSI IRQ is raised and the end of current EP's IRQ handler call but before clearing the MSI IRQ status in PCIE_MSI_INTR0_STATUS register then the new MSI IRQ is lost.

This issue has been observed in case of USB-Ethernet adapter connected to PCIe USB card running iperf3 test. iperf3 client stalls after sometime as reported here[1]. This is because XHCI raises MSI IRQ at the end of its IRQ handler but before PCIe driver has cleared its status in PCIE_MSI_INTR0_STATUS register. Hence, the new IRQ is never registered by PCIe designware core. Since, XHCI does not raise any more IRQs until the previous one is handled, it leads to a stall in communication.
Therefore driver should always clear the MSI IRQ status in PCIE_MSI_INTR0_STATUS before calling EP's IRQ handler. This make sure that DW PCIe core catches new IRQ raised during the call to EP's IRQ handler

[1] https://lkml.kernel.org/r/BN6PR18MB124994D85EAC4B5B1AD5EC56866E0@BN6PR18MB1249.namprd18.prod.outlook.com

---
 drivers/pci/dwc/pcie-designware-host.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/pci/dwc/pcie-designware-host.c +1 -1
index 8de2d5c69b1d..c29cbcd430f4 100644
@@ -68,9 +68,9 @@ irqreturn_t dw_handle_msi_irq(struct pcie_port *pp)
 		while ((pos = find_next_bit((unsigned long *) &val, 32,
 					    pos)) != 32) {
 			irq = irq_find_mapping(pp->irq_domain, i * 32 + pos);
-			generic_handle_irq(irq);
 			dw_pcie_wr_own_conf(pp, PCIE_MSI_INTR0_STATUS + i * 12,
 					    4, 1 << pos);
+			generic_handle_irq(irq);
 			pos++;
 		}
 	}
Signature · 1 line
2.16.1
Acked-by on v1 1/3
#2 lore raw

Re: [PATCH 1/3] Revert "PCI: dwc: Clear MSI interrupt status after it is handled, not before"

Quoted 3 lines · Vignesh R

Since commit 06e15e6883bed ("PCI: dwc: Clear MSI interrupt status after it is handled, not before"), MSI IRQ status in PCIE_MSI_INTR0_STATUS register is cleared after calling EP's IRQ handler.

Small nit, the SHA1 is actually:
8c934095fa2f ("PCI: dwc: Clear MSI interrupt status after it is handled, not before")

author
#3 lore raw

PATCH 2/3 PCI: dwc: pci-dra7xx: Improve MSI IRQ handling

In reply to Vignesh R

We need to ensure that there are no pending MSI IRQ vector set (i.e PCIE_MSI_INTR0_STATUS reads 0 at least once) before exiting dra7xx_pcie_msi_irq_handler(). Else, the dra7xx PCIe wrapper will not register new MSI IRQs even though PCIE_MSI_INTR0_STATUS shows IRQs are pending. Therefore, keep calling dra7xx_pcie_msi_irq_handler() until it returns IRQ_NONE, which suggests that PCIE_MSI_INTR0_STATUS is 0.

This fixes a bug, where PCIe wifi cards with 4 DMA queues like Intel 8260 used to throw following error and stall during ping/iperf3 tests.

[   97.776310] iwlwifi 0000:01:00.0: Queue 9 stuck for 2500 ms.
---
 drivers/pci/dwc/pci-dra7xx.c | 21 ++++++++++++++++++---
 1 file changed, 18 insertions(+), 3 deletions(-)
drivers/pci/dwc/pci-dra7xx.c +18 -3
index ed8558d638e5..3420cbf7b60a 100644
@@ -254,14 +254,31 @@ static irqreturn_t dra7xx_pcie_msi_irq_handler(int irq, void *arg)
 	struct dra7xx_pcie *dra7xx = arg;
 	struct dw_pcie *pci = dra7xx->pci;
 	struct pcie_port *pp = &pci->pp;
+	int count = 0;
 	unsigned long reg;
 	u32 virq, bit;
 
 	reg = dra7xx_pcie_readl(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI);
+	dra7xx_pcie_writel(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI, reg);
 
 	switch (reg) {
 	case MSI:
-		dw_handle_msi_irq(pp);
+		/*
+		 * Need to make sure no MSI IRQs are pending before
+		 * exiting handler, else the wrapper will not catch new
+		 * IRQs. So loop around till dw_handle_msi_irq() returns
+		 * IRQ_NONE
+		 */
+		while (dw_handle_msi_irq(pp) != IRQ_NONE && count < 1000)
+			count++;
+
+		if (count == 1000) {
+			dev_err(pci->dev, "too much work in msi irq\n");
+			dra7xx_pcie_writel(dra7xx,
+					   PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI,
+					   reg);
+			return IRQ_HANDLED;
+		}
 		break;
 	case INTA:
 	case INTB:
@@ -275,8 +292,6 @@ static irqreturn_t dra7xx_pcie_msi_irq_handler(int irq, void *arg)
 		break;
 	}
 
-	dra7xx_pcie_writel(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI, reg);
-
 	return IRQ_HANDLED;
 }
 
Signature · 1 line
2.16.1
#4 lore raw

Re: [PATCH 2/3] PCI: dwc: pci-dra7xx: Improve MSI IRQ handling

Quoted 40 lines · Vignesh R

We need to ensure that there are no pending MSI IRQ vector set (i.e PCIE_MSI_INTR0_STATUS reads 0 at least once) before exiting dra7xx_pcie_msi_irq_handler(). Else, the dra7xx PCIe wrapper will not register new MSI IRQs even though PCIE_MSI_INTR0_STATUS shows IRQs are pending. Therefore, keep calling dra7xx_pcie_msi_irq_handler() until it returns IRQ_NONE, which suggests that PCIE_MSI_INTR0_STATUS is 0.

This fixes a bug, where PCIe wifi cards with 4 DMA queues like Intel 8260 used to throw following error and stall during ping/iperf3 tests.

[   97.776310] iwlwifi 0000:01:00.0: Queue 9 stuck for 2500 ms.
---
 drivers/pci/dwc/pci-dra7xx.c | 21 ++++++++++++++++++---
 1 file changed, 18 insertions(+), 3 deletions(-)
drivers/pci/dwc/pci-dra7xx.c +7 -1
index ed8558d638e5..3420cbf7b60a 100644
@@ -254,14 +254,31 @@ static irqreturn_t dra7xx_pcie_msi_irq_handler(int irq, void *arg)
 	struct dra7xx_pcie *dra7xx = arg;
 	struct dw_pcie *pci = dra7xx->pci;
 	struct pcie_port *pp = &pci->pp;
+	int count = 0;
 	unsigned long reg;
 	u32 virq, bit;
 
 	reg = dra7xx_pcie_readl(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI);
+	dra7xx_pcie_writel(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI, reg);
 
 	switch (reg) {
 	case MSI:
-		dw_handle_msi_irq(pp);
+		/*
+		 * Need to make sure no MSI IRQs are pending before
+		 * exiting handler, else the wrapper will not catch new
+		 * IRQs. So loop around till dw_handle_msi_irq() returns
+		 * IRQ_NONE

drivers/pci/dwc/pci-dra7xx.c static irqreturn_t dra7xx_pcie_msi_irq_handler(int irq, void *arg) · line 280

Quoted 10 lines
drivers/pci/dwc/pci-dra7xx.c +10 -0
+		 */
+		while (dw_handle_msi_irq(pp) != IRQ_NONE && count < 1000)
+			count++;
+
+		if (count == 1000) {
+			dev_err(pci->dev, "too much work in msi irq\n");
+			dra7xx_pcie_writel(dra7xx,
+					   PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI,
+					   reg);
+			return IRQ_HANDLED;

I am not merging any code patching this IRQ handling routine anymore unless you thoroughly explain to me how this CONF_IRQSTATUS_MSI register works (and how it is related to DW registers) and why this specific host controller needs handling that is not required by any other host controller relying on dw_handle_msi_irq().

I suspect there is a code design flaw with the way this host handles IRQs and we are going to find it and fix it the way it should, not with any plaster like this patch.

Lorenzo

Quoted 16 lines
drivers/pci/dwc/pci-dra7xx.c +1 -2
+		}
 		break;
 	case INTA:
 	case INTB:
@@ -275,8 +292,6 @@ static irqreturn_t dra7xx_pcie_msi_irq_handler(int irq, void *arg)
 		break;
 	}
 
-	dra7xx_pcie_writel(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI, reg);
-
 	return IRQ_HANDLED;
 }
 
Signature · 1 line
2.16.1
author
#5 lore raw

Re: [PATCH 2/3] PCI: dwc: pci-dra7xx: Improve MSI IRQ handling

Hi,

Quoted 49 lines · Lorenzo Pieralisi
Quoted 48 lines · Vignesh R

We need to ensure that there are no pending MSI IRQ vector set (i.e PCIE_MSI_INTR0_STATUS reads 0 at least once) before exiting dra7xx_pcie_msi_irq_handler(). Else, the dra7xx PCIe wrapper will not register new MSI IRQs even though PCIE_MSI_INTR0_STATUS shows IRQs are pending. Therefore, keep calling dra7xx_pcie_msi_irq_handler() until it returns IRQ_NONE, which suggests that PCIE_MSI_INTR0_STATUS is 0.

This fixes a bug, where PCIe wifi cards with 4 DMA queues like Intel 8260 used to throw following error and stall during ping/iperf3 tests.

[   97.776310] iwlwifi 0000:01:00.0: Queue 9 stuck for 2500 ms.

---
  drivers/pci/dwc/pci-dra7xx.c | 21 ++++++++++++++++++---

  1 file changed, 18 insertions(+), 3 deletions(-)
drivers/pci/dwc/pci-dra7xx.c +1 -0
index ed8558d638e5..3420cbf7b60a 100644
@@ -254,14 +254,31 @@ static irqreturn_t dra7xx_pcie_msi_irq_handler(int irq, void *arg)
         struct dra7xx_pcie *dra7xx = arg;
         struct dw_pcie *pci = dra7xx->pci;
         struct pcie_port *pp = &pci->pp;
+     int count = 0;
         unsigned long reg;
         u32 virq, bit;


        reg = dra7xx_pcie_readl(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI); +     dra7xx_pcie_writel(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI, reg);
        switch (reg) {
        case MSI:
-             dw_handle_msi_irq(pp);
+             /*
+              * Need to make sure no MSI IRQs are pending before +              * exiting handler, else the wrapper will not catch new +              * IRQs. So loop around till dw_handle_msi_irq() returns +              * IRQ_NONE
+              */
+             while (dw_handle_msi_irq(pp) != IRQ_NONE && count < 1000) +                     count++;
+
+             if (count == 1000) {
+                     dev_err(pci->dev, "too much work in msi irq\n"); +                     dra7xx_pcie_writel(dra7xx,
+                                        PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI,

Quoted 8 lines
Quoted 2 lines

+                                        reg);
+                     return IRQ_HANDLED;

I am not merging any code patching this IRQ handling routine anymore unless you thoroughly explain to me how this CONF_IRQSTATUS_MSI register works (and how it is related to DW registers) and why this specific host controller needs handling that is not required by any other host controller relying on dw_handle_msi_irq().

Unlike other DW PCIe controllers, TI implementation has a wrapper on top of DW core. This wrapper latches the DW core level MSI and legacy interrupts and then propagates it to GIC. PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI register is present in this TI wrapper which aggregates all the MSI IRQs(PCIE_MSI_INTR0_STATUS) of DW level. They are mapped on the MSI interrupt line of PCIe controller, using a single status bit in the PCIECTRL_TI_CONF_IRQSTATUS_MSI register.

So, the irq handler, dra7xx_pcie_msi_irq_handler(), first needs to look at PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI[4] to know that its MSI IRQ and then call dw_handle_msi_irq() to handle individual MSI vectors. Driver has to make sure there are no pending vectors in DW core MSI status register before exiting handler. Otherwise next MSI IRQ will not be latched by the wrapper.

Quoted 5 lines

I suspect there is a code design flaw with the way this host handles IRQs and we are going to find it and fix it the way it should, not with any plaster like this patch.

I agree there has been some churn wrt this wrapper level IRQ handler. But, that was because hardware documentation/TRM did not match actual behavior and so it took some time to understand how the hardware is working. I have extensively tested this series on multiple problematic PCIe USB cards and PCIe WiFi cards over week long stress tests. And also had some agreement with internal hardware designers. Hardware documentations will also be updated.

Quoted 18 lines

Lorenzo

Quoted 16 lines

+             }
                break;
        case INTA:
        case INTB:

drivers/pci/dwc/pci-dra7xx.c +0 -0
@@ -275,8 +292,6 @@ static irqreturn_t dra7xx_pcie_msi_irq_handler(int irq, void *arg)
                 break;
         }


-     dra7xx_pcie_writel(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI, reg); -
        return IRQ_HANDLED;
  }

Signature · 1 line
2.16.1
Signature · 2 lines
Regards
Vignesh
author
#6 lore raw

Re: [PATCH 2/3] PCI: dwc: pci-dra7xx: Improve MSI IRQ handling

Hi Lorenzo,

Quoted 69 lines · Vignesh R

Hi,

Quoted 49 lines · Lorenzo Pieralisi
Quoted 48 lines · Vignesh R

We need to ensure that there are no pending MSI IRQ vector set (i.e PCIE_MSI_INTR0_STATUS reads 0 at least once) before exiting dra7xx_pcie_msi_irq_handler(). Else, the dra7xx PCIe wrapper will not register new MSI IRQs even though PCIE_MSI_INTR0_STATUS shows IRQs are pending. Therefore, keep calling dra7xx_pcie_msi_irq_handler() until it returns IRQ_NONE, which suggests that PCIE_MSI_INTR0_STATUS is 0.

This fixes a bug, where PCIe wifi cards with 4 DMA queues like Intel 8260 used to throw following error and stall during ping/iperf3 tests.

[   97.776310] iwlwifi 0000:01:00.0: Queue 9 stuck for 2500 ms.

---
  drivers/pci/dwc/pci-dra7xx.c | 21 ++++++++++++++++++---

  1 file changed, 18 insertions(+), 3 deletions(-)
drivers/pci/dwc/pci-dra7xx.c +1 -0
index ed8558d638e5..3420cbf7b60a 100644
@@ -254,14 +254,31 @@ static irqreturn_t dra7xx_pcie_msi_irq_handler(int irq, void *arg)
         struct dra7xx_pcie *dra7xx = arg;
         struct dw_pcie *pci = dra7xx->pci;
         struct pcie_port *pp = &pci->pp;
+     int count = 0;
         unsigned long reg;
         u32 virq, bit;


        reg = dra7xx_pcie_readl(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI); +     dra7xx_pcie_writel(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI, reg);
        switch (reg) {
        case MSI:
-             dw_handle_msi_irq(pp);
+             /*
+              * Need to make sure no MSI IRQs are pending before +              * exiting handler, else the wrapper will not catch new +              * IRQs. So loop around till dw_handle_msi_irq() returns +              * IRQ_NONE
+              */
+             while (dw_handle_msi_irq(pp) != IRQ_NONE && count < 1000) +                     count++;
+
+             if (count == 1000) {
+                     dev_err(pci->dev, "too much work in msi irq\n"); +                     dra7xx_pcie_writel(dra7xx,
+                                        PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI,

Quoted 8 lines
Quoted 2 lines

+                                        reg);
+                     return IRQ_HANDLED;

I am not merging any code patching this IRQ handling routine anymore unless you thoroughly explain to me how this CONF_IRQSTATUS_MSI register works (and how it is related to DW registers) and why this specific host controller needs handling that is not required by any other host controller relying on dw_handle_msi_irq().

Unlike other DW PCIe controllers, TI implementation has a wrapper on top of DW core. This wrapper latches the DW core level MSI and legacy interrupts and then propagates it to GIC. PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI register is present in this TI wrapper which aggregates all the MSI IRQs(PCIE_MSI_INTR0_STATUS) of DW level. They are mapped on the MSI interrupt line of PCIe controller, using a single status bit in the PCIECTRL_TI_CONF_IRQSTATUS_MSI register.

Quoted 8 lines

So, the irq handler, dra7xx_pcie_msi_irq_handler(), first needs to look at PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI[4] to know that its MSI IRQ and then call dw_handle_msi_irq() to handle individual MSI vectors. Driver has to make sure there are no pending vectors in DW core MSI status register before exiting handler. Otherwise next MSI IRQ will not be latched by the wrapper.

Did above explanation clarify CONF_IRQSTATUS_MSI register usage and the need for IRQ handler? Are you okay with this fix?

Regards
Vignesh

Quoted 34 lines
Quoted 5 lines

I suspect there is a code design flaw with the way this host handles IRQs and we are going to find it and fix it the way it should, not with any plaster like this patch.

I agree there has been some churn wrt this wrapper level IRQ handler. But, that was because hardware documentation/TRM did not match actual behavior and so it took some time to understand how the hardware is working. I have extensively tested this series on multiple problematic PCIe USB cards and PCIe WiFi cards over week long stress tests. And also had some agreement with internal hardware designers. Hardware documentations will also be updated.

Quoted 18 lines

Lorenzo

Quoted 16 lines

+             }
                break;
        case INTA:
        case INTB:

drivers/pci/dwc/pci-dra7xx.c +0 -0
@@ -275,8 +292,6 @@ static irqreturn_t dra7xx_pcie_msi_irq_handler(int irq, void *arg)
                 break;
         }


-     dra7xx_pcie_writel(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI, reg); -
        return IRQ_HANDLED;
  }

Signature · 1 line
2.16.1
#7 lore raw

Re: [PATCH 2/3] PCI: dwc: pci-dra7xx: Improve MSI IRQ handling

Quoted 81 lines · Vignesh R

Hi Lorenzo,

Quoted 69 lines · Vignesh R

Hi,

Quoted 49 lines · Lorenzo Pieralisi
Quoted 48 lines · Vignesh R

We need to ensure that there are no pending MSI IRQ vector set (i.e PCIE_MSI_INTR0_STATUS reads 0 at least once) before exiting dra7xx_pcie_msi_irq_handler(). Else, the dra7xx PCIe wrapper will not register new MSI IRQs even though PCIE_MSI_INTR0_STATUS shows IRQs are pending. Therefore, keep calling dra7xx_pcie_msi_irq_handler() until it returns IRQ_NONE, which suggests that PCIE_MSI_INTR0_STATUS is 0.

This fixes a bug, where PCIe wifi cards with 4 DMA queues like Intel 8260 used to throw following error and stall during ping/iperf3 tests.

[   97.776310] iwlwifi 0000:01:00.0: Queue 9 stuck for 2500 ms.

---
  drivers/pci/dwc/pci-dra7xx.c | 21 ++++++++++++++++++---

  1 file changed, 18 insertions(+), 3 deletions(-)
drivers/pci/dwc/pci-dra7xx.c +1 -0
index ed8558d638e5..3420cbf7b60a 100644
@@ -254,14 +254,31 @@ static irqreturn_t dra7xx_pcie_msi_irq_handler(int irq, void *arg)
         struct dra7xx_pcie *dra7xx = arg;
         struct dw_pcie *pci = dra7xx->pci;
         struct pcie_port *pp = &pci->pp;
+     int count = 0;
         unsigned long reg;
         u32 virq, bit;


        reg = dra7xx_pcie_readl(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI); +     dra7xx_pcie_writel(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI, reg);
        switch (reg) {
        case MSI:
-             dw_handle_msi_irq(pp);
+             /*
+              * Need to make sure no MSI IRQs are pending before +              * exiting handler, else the wrapper will not catch new +              * IRQs. So loop around till dw_handle_msi_irq() returns +              * IRQ_NONE
+              */
+             while (dw_handle_msi_irq(pp) != IRQ_NONE && count < 1000) +                     count++;
+
+             if (count == 1000) {
+                     dev_err(pci->dev, "too much work in msi irq\n"); +                     dra7xx_pcie_writel(dra7xx,
+                                        PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI,

Quoted 8 lines
Quoted 2 lines

+                                        reg);
+                     return IRQ_HANDLED;

I am not merging any code patching this IRQ handling routine anymore unless you thoroughly explain to me how this CONF_IRQSTATUS_MSI register works (and how it is related to DW registers) and why this specific host controller needs handling that is not required by any other host controller relying on dw_handle_msi_irq().

Unlike other DW PCIe controllers, TI implementation has a wrapper on top of DW core. This wrapper latches the DW core level MSI and legacy interrupts and then propagates it to GIC. PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI register is present in this TI wrapper which aggregates all the MSI IRQs(PCIE_MSI_INTR0_STATUS) of DW level. They are mapped on the MSI interrupt line of PCIe controller, using a single status bit in the PCIECTRL_TI_CONF_IRQSTATUS_MSI register.

Quoted 8 lines

So, the irq handler, dra7xx_pcie_msi_irq_handler(), first needs to look at PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI[4] to know that its MSI IRQ and then call dw_handle_msi_irq() to handle individual MSI vectors. Driver has to make sure there are no pending vectors in DW core MSI status register before exiting handler. Otherwise next MSI IRQ will not be latched by the wrapper.

Quoted 2 lines

Did above explanation clarify CONF_IRQSTATUS_MSI register usage and the need for IRQ handler? Are you okay with this fix?

Hi Vignesh,

I will get back to it shortly, thanks for your patience.

Lorenzo

Quoted 37 lines

Regards
Vignesh

Quoted 34 lines
Quoted 5 lines

I suspect there is a code design flaw with the way this host handles IRQs and we are going to find it and fix it the way it should, not with any plaster like this patch.

I agree there has been some churn wrt this wrapper level IRQ handler. But, that was because hardware documentation/TRM did not match actual behavior and so it took some time to understand how the hardware is working. I have extensively tested this series on multiple problematic PCIe USB cards and PCIe WiFi cards over week long stress tests. And also had some agreement with internal hardware designers. Hardware documentations will also be updated.

Quoted 18 lines

Lorenzo

Quoted 16 lines

+             }
                break;
        case INTA:
        case INTB:

drivers/pci/dwc/pci-dra7xx.c +0 -0
@@ -275,8 +292,6 @@ static irqreturn_t dra7xx_pcie_msi_irq_handler(int irq, void *arg)
                 break;
         }


-     dra7xx_pcie_writel(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI, reg); -
        return IRQ_HANDLED;
  }

Signature · 1 line
2.16.1
#8 lore raw

Re: [PATCH 2/3] PCI: dwc: pci-dra7xx: Improve MSI IRQ handling

In reply to Vignesh R
Quoted 69 lines · Vignesh R

Hi,

Quoted 49 lines · Lorenzo Pieralisi
Quoted 48 lines · Vignesh R

We need to ensure that there are no pending MSI IRQ vector set (i.e PCIE_MSI_INTR0_STATUS reads 0 at least once) before exiting dra7xx_pcie_msi_irq_handler(). Else, the dra7xx PCIe wrapper will not register new MSI IRQs even though PCIE_MSI_INTR0_STATUS shows IRQs are pending. Therefore, keep calling dra7xx_pcie_msi_irq_handler() until it returns IRQ_NONE, which suggests that PCIE_MSI_INTR0_STATUS is 0.

This fixes a bug, where PCIe wifi cards with 4 DMA queues like Intel 8260 used to throw following error and stall during ping/iperf3 tests.

[   97.776310] iwlwifi 0000:01:00.0: Queue 9 stuck for 2500 ms.

---
  drivers/pci/dwc/pci-dra7xx.c | 21 ++++++++++++++++++---

  1 file changed, 18 insertions(+), 3 deletions(-)
drivers/pci/dwc/pci-dra7xx.c +1 -0
index ed8558d638e5..3420cbf7b60a 100644
@@ -254,14 +254,31 @@ static irqreturn_t dra7xx_pcie_msi_irq_handler(int irq, void *arg)
         struct dra7xx_pcie *dra7xx = arg;
         struct dw_pcie *pci = dra7xx->pci;
         struct pcie_port *pp = &pci->pp;
+     int count = 0;
         unsigned long reg;
         u32 virq, bit;


        reg = dra7xx_pcie_readl(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI); +     dra7xx_pcie_writel(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI, reg);
        switch (reg) {
        case MSI:
-             dw_handle_msi_irq(pp);
+             /*
+              * Need to make sure no MSI IRQs are pending before +              * exiting handler, else the wrapper will not catch new +              * IRQs. So loop around till dw_handle_msi_irq() returns +              * IRQ_NONE
+              */
+             while (dw_handle_msi_irq(pp) != IRQ_NONE && count < 1000) +                     count++;
+
+             if (count == 1000) {
+                     dev_err(pci->dev, "too much work in msi irq\n"); +                     dra7xx_pcie_writel(dra7xx,
+                                        PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI,

Quoted 8 lines
Quoted 2 lines

+                                        reg);
+                     return IRQ_HANDLED;

I am not merging any code patching this IRQ handling routine anymore unless you thoroughly explain to me how this CONF_IRQSTATUS_MSI register works (and how it is related to DW registers) and why this specific host controller needs handling that is not required by any other host controller relying on dw_handle_msi_irq().

Unlike other DW PCIe controllers, TI implementation has a wrapper on top of DW core. This wrapper latches the DW core level MSI and legacy interrupts and then propagates it to GIC. PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI register is present in this TI wrapper which aggregates all the MSI IRQs(PCIE_MSI_INTR0_STATUS) of DW level. They are mapped on the MSI interrupt line of PCIe controller, using a single status bit in the PCIECTRL_TI_CONF_IRQSTATUS_MSI register.

Quoted 4 lines

So, the irq handler, dra7xx_pcie_msi_irq_handler(), first needs to look at PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI[4] to know that its MSI IRQ and then call dw_handle_msi_irq() to handle individual MSI vectors. Driver has to make sure there are no pending vectors in DW core MSI

How can it make *sure* ? And what makes the wrapper latch MSI IRQs again ?

Quoted 2 lines

status register before exiting handler. Otherwise next MSI IRQ will not be latched by the wrapper.

I am sorry but I do not understand how this works - what is the condition that makes wrapper latch IRQs again ? This is at least racy, if not outright broken.

That count == 1000 is a symptom there is something broken on how this driver handles IRQs and I have the impression that we are applying plasters on top of plasters to make it less broken than it actually is.

Quoted 9 lines
Quoted 4 lines

I suspect there is a code design flaw with the way this host handles IRQs and we are going to find it and fix it the way it should, not with any plaster like this patch.

I agree there has been some churn wrt this wrapper level IRQ handler. But, that was because hardware documentation/TRM did not match actual behavior and so it took some time to understand how the hardware is working.

How does HW work :) ? Please explain in detail how this works in HW then we will get to the code.

Thanks,
Lorenzo

Quoted 28 lines

I have extensively tested this series on multiple problematic PCIe USB cards and PCIe WiFi cards over week long stress tests. And also had some agreement with internal hardware designers. Hardware documentations will also be updated.

Quoted 18 lines

Lorenzo

Quoted 16 lines

+             }
                break;
        case INTA:
        case INTB:

drivers/pci/dwc/pci-dra7xx.c +0 -0
@@ -275,8 +292,6 @@ static irqreturn_t dra7xx_pcie_msi_irq_handler(int irq, void *arg)
                 break;
         }


-     dra7xx_pcie_writel(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI, reg); -
        return IRQ_HANDLED;
  }

Signature · 1 line
2.16.1
Signature · 2 lines
Regards
Vignesh
author
#9 lore raw

Re: [PATCH 2/3] PCI: dwc: pci-dra7xx: Improve MSI IRQ handling

Quoted 70 lines · Lorenzo Pieralisi
Quoted 69 lines · Vignesh R

Hi,

Quoted 49 lines · Lorenzo Pieralisi
Quoted 48 lines · Vignesh R

We need to ensure that there are no pending MSI IRQ vector set (i.e PCIE_MSI_INTR0_STATUS reads 0 at least once) before exiting dra7xx_pcie_msi_irq_handler(). Else, the dra7xx PCIe wrapper will not register new MSI IRQs even though PCIE_MSI_INTR0_STATUS shows IRQs are pending. Therefore, keep calling dra7xx_pcie_msi_irq_handler() until it returns IRQ_NONE, which suggests that PCIE_MSI_INTR0_STATUS is 0.

This fixes a bug, where PCIe wifi cards with 4 DMA queues like Intel 8260 used to throw following error and stall during ping/iperf3 tests.

[   97.776310] iwlwifi 0000:01:00.0: Queue 9 stuck for 2500 ms.

---
  drivers/pci/dwc/pci-dra7xx.c | 21 ++++++++++++++++++---

  1 file changed, 18 insertions(+), 3 deletions(-)
drivers/pci/dwc/pci-dra7xx.c +1 -0
index ed8558d638e5..3420cbf7b60a 100644
@@ -254,14 +254,31 @@ static irqreturn_t dra7xx_pcie_msi_irq_handler(int irq, void *arg)
         struct dra7xx_pcie *dra7xx = arg;
         struct dw_pcie *pci = dra7xx->pci;
         struct pcie_port *pp = &pci->pp;
+     int count = 0;
         unsigned long reg;
         u32 virq, bit;


        reg = dra7xx_pcie_readl(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI); +     dra7xx_pcie_writel(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI, reg);
        switch (reg) {
        case MSI:
-             dw_handle_msi_irq(pp);
+             /*
+              * Need to make sure no MSI IRQs are pending before +              * exiting handler, else the wrapper will not catch new +              * IRQs. So loop around till dw_handle_msi_irq() returns +              * IRQ_NONE
+              */
+             while (dw_handle_msi_irq(pp) != IRQ_NONE && count < 1000) +                     count++;
+
+             if (count == 1000) {
+                     dev_err(pci->dev, "too much work in msi irq\n"); +                     dra7xx_pcie_writel(dra7xx,
+                                        PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI,

Quoted 8 lines
Quoted 2 lines

+                                        reg);
+                     return IRQ_HANDLED;

I am not merging any code patching this IRQ handling routine anymore unless you thoroughly explain to me how this CONF_IRQSTATUS_MSI register works (and how it is related to DW registers) and why this specific host controller needs handling that is not required by any other host controller relying on dw_handle_msi_irq().

Unlike other DW PCIe controllers, TI implementation has a wrapper on top of DW core. This wrapper latches the DW core level MSI and legacy interrupts and then propagates it to GIC. PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI register is present in this TI wrapper which aggregates all the MSI IRQs(PCIE_MSI_INTR0_STATUS) of DW level. They are mapped on the MSI interrupt line of PCIe controller, using a single status bit in the PCIECTRL_TI_CONF_IRQSTATUS_MSI register.

Quoted 8 lines
Quoted 4 lines

So, the irq handler, dra7xx_pcie_msi_irq_handler(), first needs to look at PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI[4] to know that its MSI IRQ and then call dw_handle_msi_irq() to handle individual MSI vectors. Driver has to make sure there are no pending vectors in DW core MSI

How can it make *sure* ? And what makes the wrapper latch MSI IRQs again ?

This is the sequence that I got from discussion with internal HW team:
1. read CONF_IRQSTATUS_MSI in wrapper and check if MSI bit
2. clear CONF_IRQSTATUS_MSI
3. read, clear and handle PCIE_MSI_INTR0_STATUS vectors
4. repeat step 3 until PCIE_MSI_INTR0_STATUS reads 0

If read of PCIE_MSI_INTR0_STATUS returns 0 at least once, then its guaranteed that the next time any vector is set in PCIE_MSI_INTR0_STATUS register(due to MSI IRQ), wrapper will latch it and raise an IRQ to CPU.

Quoted 11 lines
Quoted 2 lines

status register before exiting handler. Otherwise next MSI IRQ will not be latched by the wrapper.

I am sorry but I do not understand how this works - what is the condition that makes wrapper latch IRQs again ? This is at least racy, if not outright broken.

That count == 1000 is a symptom there is something broken on how this driver handles IRQs and I have the impression that we are applying plasters on top of plasters to make it less broken than it actually is.

It is an upper bound on how many times driver looks at PCIE_MSI_INTR0_STATUS register, so that there is no infinite looping when there is an IRQ flood due to misbehaving EP. count == 1000 condition should not happen and it means something is wrong in the system. I haven't hit this situation in testing
I can either remove this or put a WARN_ON to say this situation should not have happened, if that makes you more comfortable with the patch.

Quoted 5 lines
Quoted 5 lines
Quoted 4 lines

I suspect there is a code design flaw with the way this host handles IRQs and we are going to find it and fix it the way it should, not with any plaster like this patch.

Quoted 8 lines
Quoted 4 lines

I agree there has been some churn wrt this wrapper level IRQ handler. But, that was because hardware documentation/TRM did not match actual behavior and so it took some time to understand how the hardware is working.

How does HW work :) ? Please explain in detail how this works in HW then we will get to the code.

Software needs to ensure that PCIE_MSI_INTR0_STATUS needs be 0 by reading it. Then, when the next MSI IRQ is raised CONF_IRQSTATUS_MSI register in the wrapper will latch the next IRQ.

This is my current knowledge, let me know if you need to know anything specifically, I will try to ask HW team.

Regards
Vignesh

Quoted 31 lines

Thanks,
Lorenzo

Quoted 28 lines

I have extensively tested this series on multiple problematic PCIe USB cards and PCIe WiFi cards over week long stress tests. And also had some agreement with internal hardware designers. Hardware documentations will also be updated.

Quoted 18 lines

Lorenzo

Quoted 16 lines

+             }
                break;
        case INTA:
        case INTB:

drivers/pci/dwc/pci-dra7xx.c +0 -0
@@ -275,8 +292,6 @@ static irqreturn_t dra7xx_pcie_msi_irq_handler(int irq, void *arg)
                 break;
         }


-     dra7xx_pcie_writel(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI, reg); -
        return IRQ_HANDLED;
  }

Signature · 1 line
2.16.1
Signature · 2 lines
Regards
Vignesh
#10 lore raw

Re: [PATCH 2/3] PCI: dwc: pci-dra7xx: Improve MSI IRQ handling

[CC'ed Gustavo]

Quoted 88 lines · Vignesh R
Quoted 70 lines · Lorenzo Pieralisi
Quoted 69 lines · Vignesh R

Hi,

Quoted 49 lines · Lorenzo Pieralisi
Quoted 48 lines · Vignesh R

We need to ensure that there are no pending MSI IRQ vector set (i.e PCIE_MSI_INTR0_STATUS reads 0 at least once) before exiting dra7xx_pcie_msi_irq_handler(). Else, the dra7xx PCIe wrapper will not register new MSI IRQs even though PCIE_MSI_INTR0_STATUS shows IRQs are pending. Therefore, keep calling dra7xx_pcie_msi_irq_handler() until it returns IRQ_NONE, which suggests that PCIE_MSI_INTR0_STATUS is 0.

This fixes a bug, where PCIe wifi cards with 4 DMA queues like Intel 8260 used to throw following error and stall during ping/iperf3 tests.

[   97.776310] iwlwifi 0000:01:00.0: Queue 9 stuck for 2500 ms.

---
  drivers/pci/dwc/pci-dra7xx.c | 21 ++++++++++++++++++---

  1 file changed, 18 insertions(+), 3 deletions(-)
drivers/pci/dwc/pci-dra7xx.c +1 -0
index ed8558d638e5..3420cbf7b60a 100644
@@ -254,14 +254,31 @@ static irqreturn_t dra7xx_pcie_msi_irq_handler(int irq, void *arg)
         struct dra7xx_pcie *dra7xx = arg;
         struct dw_pcie *pci = dra7xx->pci;
         struct pcie_port *pp = &pci->pp;
+     int count = 0;
         unsigned long reg;
         u32 virq, bit;


        reg = dra7xx_pcie_readl(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI); +     dra7xx_pcie_writel(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI, reg);
        switch (reg) {
        case MSI:
-             dw_handle_msi_irq(pp);
+             /*
+              * Need to make sure no MSI IRQs are pending before +              * exiting handler, else the wrapper will not catch new +              * IRQs. So loop around till dw_handle_msi_irq() returns +              * IRQ_NONE
+              */
+             while (dw_handle_msi_irq(pp) != IRQ_NONE && count < 1000) +                     count++;
+
+             if (count == 1000) {
+                     dev_err(pci->dev, "too much work in msi irq\n"); +                     dra7xx_pcie_writel(dra7xx,
+                                        PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI,

Quoted 8 lines
Quoted 2 lines

+                                        reg);
+                     return IRQ_HANDLED;

I am not merging any code patching this IRQ handling routine anymore unless you thoroughly explain to me how this CONF_IRQSTATUS_MSI register works (and how it is related to DW registers) and why this specific host controller needs handling that is not required by any other host controller relying on dw_handle_msi_irq().

Unlike other DW PCIe controllers, TI implementation has a wrapper on top of DW core. This wrapper latches the DW core level MSI and legacy interrupts and then propagates it to GIC. PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI register is present in this TI wrapper which aggregates all the MSI IRQs(PCIE_MSI_INTR0_STATUS) of DW level. They are mapped on the MSI interrupt line of PCIe controller, using a single status bit in the PCIECTRL_TI_CONF_IRQSTATUS_MSI register.

Quoted 8 lines
Quoted 4 lines

So, the irq handler, dra7xx_pcie_msi_irq_handler(), first needs to look at PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI[4] to know that its MSI IRQ and then call dw_handle_msi_irq() to handle individual MSI vectors. Driver has to make sure there are no pending vectors in DW core MSI

How can it make *sure* ? And what makes the wrapper latch MSI IRQs again ?

This is the sequence that I got from discussion with internal HW team:
1. read CONF_IRQSTATUS_MSI in wrapper and check if MSI bit
2. clear CONF_IRQSTATUS_MSI
3. read, clear and handle PCIE_MSI_INTR0_STATUS vectors
4. repeat step 3 until PCIE_MSI_INTR0_STATUS reads 0

Quoted 7 lines

If read of PCIE_MSI_INTR0_STATUS returns 0 at least once, then its guaranteed that the next time any vector is set in PCIE_MSI_INTR0_STATUS register(due to MSI IRQ), wrapper will latch it and raise an IRQ to CPU.

Quoted 2 lines
Quoted 2 lines

status register before exiting handler. Otherwise next MSI IRQ will not be latched by the wrapper.

Hi Vignesh,

thank you for explaining. To me - this looks like it is better to write a separate IRQ chip implementation for DRA7XX, let's face it this patch is a bit hackish (and will prevent us from removing dw_handle_msi_irq() altogether - which is something we want to achieve) and it is clutching at straws to re-use DWC MSI code but honestly it is a bit of stretch.

This would duplicate some code, yes but on the other hand it would give you full control of the (DWC) HW.

You can have a look at pcie-tango.c or the mobiveil patches I have just reviewed:

https://patchwork.ozlabs.org/patch/878494/

Please note Gustavo's reworked DWC MSI handling that is now queued for v4.17 in my pci/dwc-msi branch - please do have a look at it too (and if Gustavo has comments they are welcome).

I do not like the current implementation - sorry, we have to come up with something cleaner, I do not think that's so complicated.

I do not mind reviewing intermediate patches as long as we get to a cleaner solution.

Thanks,
Lorenzo

Quoted 74 lines
Quoted 9 lines

I am sorry but I do not understand how this works - what is the condition that makes wrapper latch IRQs again ? This is at least racy, if not outright broken.

That count == 1000 is a symptom there is something broken on how this driver handles IRQs and I have the impression that we are applying plasters on top of plasters to make it less broken than it actually is.

It is an upper bound on how many times driver looks at PCIE_MSI_INTR0_STATUS register, so that there is no infinite looping when there is an IRQ flood due to misbehaving EP. count == 1000 condition should not happen and it means something is wrong in the system. I haven't hit this situation in testing
I can either remove this or put a WARN_ON to say this situation should not have happened, if that makes you more comfortable with the patch.

Quoted 5 lines
Quoted 5 lines
Quoted 4 lines

I suspect there is a code design flaw with the way this host handles IRQs and we are going to find it and fix it the way it should, not with any plaster like this patch.

Quoted 8 lines
Quoted 4 lines

I agree there has been some churn wrt this wrapper level IRQ handler. But, that was because hardware documentation/TRM did not match actual behavior and so it took some time to understand how the hardware is working.

How does HW work :) ? Please explain in detail how this works in HW then we will get to the code.

Software needs to ensure that PCIE_MSI_INTR0_STATUS needs be 0 by reading it. Then, when the next MSI IRQ is raised CONF_IRQSTATUS_MSI register in the wrapper will latch the next IRQ.

This is my current knowledge, let me know if you need to know anything specifically, I will try to ask HW team.

Regards
Vignesh

Quoted 31 lines

Thanks,
Lorenzo

Quoted 28 lines

I have extensively tested this series on multiple problematic PCIe USB cards and PCIe WiFi cards over week long stress tests. And also had some agreement with internal hardware designers. Hardware documentations will also be updated.

Quoted 18 lines

Lorenzo

Quoted 16 lines

+             }
                break;
        case INTA:
        case INTB:

drivers/pci/dwc/pci-dra7xx.c +0 -0
@@ -275,8 +292,6 @@ static irqreturn_t dra7xx_pcie_msi_irq_handler(int irq, void *arg)
                 break;
         }


-     dra7xx_pcie_writel(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI, reg); -
        return IRQ_HANDLED;
  }

Signature · 1 line
2.16.1
Signature · 2 lines
Regards
Vignesh
author
#11 lore raw

PATCH 3/3 PCI: dwc: pci-dra7xx: Handle legacy and MSI IRQs together

In reply to Vignesh R

Currently, pci-dra7xx driver handles MSI and Legacy IRQs exclusive of each other. This is not true, as there maybe both legacy and MSI IRQs raised at the same time. Fix this by making sure that driver handles both MSI and legacy IRQs in the IRQ handler.

---
 drivers/pci/dwc/pci-dra7xx.c | 11 +++--------
 1 file changed, 3 insertions(+), 8 deletions(-)
drivers/pci/dwc/pci-dra7xx.c +3 -8
index 3420cbf7b60a..1ddb62620acc 100644
@@ -261,8 +261,7 @@ static irqreturn_t dra7xx_pcie_msi_irq_handler(int irq, void *arg)
 	reg = dra7xx_pcie_readl(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI);
 	dra7xx_pcie_writel(dra7xx, PCIECTRL_DRA7XX_CONF_IRQSTATUS_MSI, reg);
 
-	switch (reg) {
-	case MSI:
+	if (reg & MSI) {
 		/*
 		 * Need to make sure no MSI IRQs are pending before
 		 * exiting handler, else the wrapper will not catch new
@@ -279,17 +278,13 @@ static irqreturn_t dra7xx_pcie_msi_irq_handler(int irq, void *arg)
 					   reg);
 			return IRQ_HANDLED;
 		}
-		break;
-	case INTA:
-	case INTB:
-	case INTC:
-	case INTD:
+	}
+	if (reg & LEG_EP_INTERRUPTS) {
 		for_each_set_bit(bit, &reg, PCI_NUM_INTX) {
 			virq = irq_find_mapping(dra7xx->irq_domain, bit);
 			if (virq)
 				generic_handle_irq(virq);
 		}
-		break;
 	}
 
 	return IRQ_HANDLED;
Signature · 1 line
2.16.1