PCI: fixup PCI_INTERRUPT_LINE for VMD downstream devices
PATCH 1/1 PCI: fixup PCI_INTERRUPT_LINE for VMD downstream devices
VMD does not support legacy interrupts for devices downstream from a VMD endpoint. So initialize the PCI_INTERRUPT_LINE to 0 for these devices to ensure we don't try to set up a legacy irq for them.
Note: This patch was proposed by Jim, I am trying to upstream it.
--- arch/x86/pci/fixup.c | 14 ++++++++++++++ 1 file changed, 14 insertions(+)
arch/x86/pci/fixup.c +14 -0
@@ -653,6 +653,20 @@ static void quirk_no_aersid(struct pci_dev *pdev)DECLARE_PCI_FIXUP_CLASS_EARLY (PCI_VENDOR_ID_INTEL ,PCI_ANY_ID ,PCI_CLASS_BRIDGE_PCI ,8 , quirk_no_aersid); +#if IS_ENABLED (CONFIG_VMD ) +/* +* VMD does not support legacy interrupts for downstream devices. +* So PCI_INTERRPUT_LINE needs to be initialized to 0 to ensure OS +* doesn't try to configure a legacy irq. +*/ +static void quirk_vmd_interrupt_line (struct pci_dev *dev) +{ +if (is_vmd (dev->bus)) +pci_write_config_byte (dev,PCI_INTERRUPT_LINE ,0 ); +} +DECLARE_PCI_FIXUP_HEADER (PCI_ANY_ID ,PCI_ANY_ID , quirk_vmd_interrupt_line); +#endif +static void quirk_intel_th_dnv (struct pci_dev *dev) {struct resource *r = &dev->resource[4 ];
Signature · 1 line
2.39.1
arch/x86/pci/fixup.c static void quirk_no_aersid(struct pci_dev *pdev) · line 661
Quoted 6 lines · Nirmal Patel
arch/x86/pci/fixup.c +6
-0+#if IS_ENABLED (CONFIG_VMD ) +/* +* VMD does not support legacy interrupts for downstream devices. +* So PCI_INTERRPUT_LINE needs to be initialized to 0 to ensure OS +* doesn't try to configure a legacy irq. +*/
The wording is a bit weird, Linux is the OS normally. Or is this about guest OSes in virtualized environments? Given how VMD bypasses a lot of boundaries can we even assign individual devices to VFIO? And if so is that actually safe or should we prohibit it?
Quoted 13 lines · On Wed, 24 Jul 2024 11:36:31 -0700 Christoph Hellwig <[email protected]> wrote:
arch/x86/pci/fixup.c static void quirk_no_aersid(struct pci_dev *pdev) · line 661
Quoted 6 lines · Nirmal Patel
arch/x86/pci/fixup.c +6
-0+#if IS_ENABLED (CONFIG_VMD ) +/* +* VMD does not support legacy interrupts for downstream devices. +* So PCI_INTERRPUT_LINE needs to be initialized to 0 to ensure OS +* doesn't try to configure a legacy irq. +*/ The wording is a bit weird, Linux is the OS normally. Or is this about guest OSes in virtualized environments? Given how VMD bypasses a lot of boundaries can we even assign individual devices to VFIO? And if so is that actually safe or should we prohibit it?
This is for host OS. I will add better description as suggested by Mani.
Quoted 3 lines · Nirmal Patel
VMD does not support legacy interrupts for devices downstream from a VMD endpoint. So initialize the
PCI_INTERRUPT_LINEto 0 for these devices to ensure we don't try to set up a legacy irq for them.
s/legacy interrupts/INTx/
s/legacy irq/INTx/
Quoted 13 lines
Note: This patch was proposed by Jim, I am trying to upstream it.
Signed-off-by Jim Harris <[email protected]>Signed-off-by Nirmal Patel <[email protected]>--- arch/x86/pci/fixup.c | 14 ++++++++++++++ 1 file changed, 14 insertions(+)arch/x86/pci/fixup.c +0
-0
arch/x86/pci/fixup.c static void quirk_no_aersid(struct pci_dev *pdev) · line 660
Quoted 8 lines
arch/x86/pci/fixup.c +5
-0@@ -653,6 +653,20 @@ static void quirk_no_aersid(struct pci_dev *pdev)DECLARE_PCI_FIXUP_CLASS_EARLY (PCI_VENDOR_ID_INTEL ,PCI_ANY_ID ,PCI_CLASS_BRIDGE_PCI ,8 , quirk_no_aersid); +#if IS_ENABLED (CONFIG_VMD ) +/* +* VMD does not support legacy interrupts for downstream devices. +* So PCI_INTERRPUT_LINE needs to be initialized to 0 to ensure OS +* doesn't try to configure a legacy irq.
s/legacy interrupts/INTx/
s/PCI_INTERRPUT_LINE/PCI_INTERRUPT_LINE/
arch/x86/pci/fixup.c static void quirk_no_aersid(struct pci_dev *pdev) · line 667
Quoted 7 lines
arch/x86/pci/fixup.c +7
-0+*/ +static void quirk_vmd_interrupt_line (struct pci_dev *dev) +{ +if (is_vmd (dev->bus)) +pci_write_config_byte (dev,PCI_INTERRUPT_LINE ,0 ); +} +DECLARE_PCI_FIXUP_HEADER (PCI_ANY_ID ,PCI_ANY_ID , quirk_vmd_interrupt_line);
A quirk for every PCI device, even on systems without VMD, seems like kind of a clumsy way to deal with this.
Conceptually, I would expect a host bridge driver (VMD acts like a host bridge in this case) to know whether it supports INTx, and if the driver knows it doesn't support INTx or it has no _PRT or DT description of INTx routing to use, an attempt to configure INTx should just fail naturally.
I don't claim this is how host bridge drivers actually work; I just think it's the way they *should* work.
Quoted 8 lines
arch/x86/pci/fixup.c +2
-0+#endif +static void quirk_intel_th_dnv (struct pci_dev *dev) {struct resource *r = &dev->resource[4 ];Signature · 1 line
2.39.1
Quoted 44 lines · Bjorn Helgaas
Quoted 3 lines · Nirmal Patel
VMD does not support legacy interrupts for devices downstream from a VMD endpoint. So initialize the
PCI_INTERRUPT_LINEto 0 for these devices to ensure we don't try to set up a legacy irq for them.s/legacy interrupts/INTx/
s/legacy irq/INTx/Quoted 13 lines
Note: This patch was proposed by Jim, I am trying to upstream it.
Signed-off-by Jim Harris <[email protected]>Signed-off-by Nirmal Patel <[email protected]>--- arch/x86/pci/fixup.c | 14 ++++++++++++++ 1 file changed, 14 insertions(+)arch/x86/pci/fixup.c +0
-0arch/x86/pci/fixup.c static void quirk_no_aersid(struct pci_dev *pdev) · line 660
Quoted 8 lines
arch/x86/pci/fixup.c +5
-0@@ -653,6 +653,20 @@ static void quirk_no_aersid(struct pci_dev *pdev)DECLARE_PCI_FIXUP_CLASS_EARLY (PCI_VENDOR_ID_INTEL ,PCI_ANY_ID ,PCI_CLASS_BRIDGE_PCI ,8 , quirk_no_aersid); +#if IS_ENABLED (CONFIG_VMD ) +/* +* VMD does not support legacy interrupts for downstream devices. +* So PCI_INTERRPUT_LINE needs to be initialized to 0 to ensure OS +* doesn't try to configure a legacy irq. s/legacy interrupts/INTx/
s/PCI_INTERRPUT_LINE/PCI_INTERRUPT_LINE/arch/x86/pci/fixup.c static void quirk_no_aersid(struct pci_dev *pdev) · line 667
Quoted 7 lines
arch/x86/pci/fixup.c +7
-0+*/ +static void quirk_vmd_interrupt_line (struct pci_dev *dev) +{ +if (is_vmd (dev->bus)) +pci_write_config_byte (dev,PCI_INTERRUPT_LINE ,0 ); +} +DECLARE_PCI_FIXUP_HEADER (PCI_ANY_ID ,PCI_ANY_ID , quirk_vmd_interrupt_line);A quirk for every PCI device, even on systems without VMD, seems like kind of a clumsy way to deal with this.
Quoted 9 lines
Conceptually, I would expect a host bridge driver (VMD acts like a host bridge in this case) to know whether it supports INTx, and if the driver knows it doesn't support INTx or it has no
_PRTor DT description of INTx routing to use, an attempt to configure INTx should just fail naturally.I don't claim this is how host bridge drivers actually work; I just think it's the way they *should* work.
Absolutely! This patch is fixing the issue in a wrong place. There are existing DT based host bridge drivers that disable INTx due to lack of hardware capability. The driver just need to nullify pci_host_bridge::map_irq callback.
- Mani
Signature · 1 line
மணிவண்ணன் சதாசிவம்
Quoted 58 lines · On Thu, 25 Jul 2024 09:40:13 +0530 Manivannan Sadhasivam…
Quoted 47 lines · Bjorn Helgaas
Quoted 4 lines · Nirmal Patel
VMD does not support legacy interrupts for devices downstream from a VMD endpoint. So initialize the
PCI_INTERRUPT_LINEto 0 for these devices to ensure we don't try to set up a legacy irq for them.s/legacy interrupts/INTx/
s/legacy irq/INTx/Quoted 20 lines
Note: This patch was proposed by Jim, I am trying to upstream it.
Signed-off-by Jim Harris <[email protected]>Signed-off-by Nirmal Patel <[email protected]>--- arch/x86/pci/fixup.c | 14 ++++++++++++++ 1 file changed, 14 insertions(+)arch/x86/pci/fixup.c +4
-0@@ -653,6 +653,20 @@ static void quirk_no_aersid(struct pci_dev *pdev) DECLARE_PCI_FIXUP_CLASS_EARLY(PCI_VENDOR_ID_INTEL,PCI_ANY_ID ,PCI_CLASS_BRIDGE_PCI ,8 , quirk_no_aersid); +#if IS_ENABLED (CONFIG_VMD ) +/* +* VMD does not support legacy interrupts for downstream devices. +* So PCI_INTERRPUT_LINE needs to be initialized to 0 to ensure Quoted 2 lines
OS
+ * doesn't try to configure a legacy irq.s/legacy interrupts/INTx/
s/PCI_INTERRPUT_LINE/PCI_INTERRUPT_LINE/Quoted 8 lines
+ */
arch/x86/pci/fixup.c +4
-0+static void quirk_vmd_interrupt_line (struct pci_dev *dev) +{ +if (is_vmd (dev->bus)) + pci_write_config_byte(dev, PCI_INTERRUPT_LINE,0); +}
+DECLARE_PCI_FIXUP_HEADER(PCI_ANY_ID,PCI_ANY_ID,quirk_vmd_interrupt_line);A quirk for every PCI device, even on systems without VMD, seems like kind of a clumsy way to deal with this.
Quoted 9 lines
Conceptually, I would expect a host bridge driver (VMD acts like a host bridge in this case) to know whether it supports INTx, and if the driver knows it doesn't support INTx or it has no
_PRTor DT description of INTx routing to use, an attempt to configure INTx should just fail naturally.I don't claim this is how host bridge drivers actually work; I just think it's the way they *should* work.
Quoted 7 lines
Absolutely! This patch is fixing the issue in a wrong place. There are existing DT based host bridge drivers that disable INTx due to lack of hardware capability. The driver just need to nullify
pci_host_bridge::map_irqcallback.- Mani
Thanks for the response and suggestion. I will make adjustments in VMD driver code.
-nirmal
Quoted 58 lines · On Thu, 25 Jul 2024 09:40:13 +0530 Manivannan Sadhasivam…
Quoted 47 lines · Bjorn Helgaas
Quoted 4 lines · Nirmal Patel
VMD does not support legacy interrupts for devices downstream from a VMD endpoint. So initialize the
PCI_INTERRUPT_LINEto 0 for these devices to ensure we don't try to set up a legacy irq for them.s/legacy interrupts/INTx/
s/legacy irq/INTx/Quoted 20 lines
Note: This patch was proposed by Jim, I am trying to upstream it.
Signed-off-by Jim Harris <[email protected]>Signed-off-by Nirmal Patel <[email protected]>--- arch/x86/pci/fixup.c | 14 ++++++++++++++ 1 file changed, 14 insertions(+)arch/x86/pci/fixup.c +4
-0@@ -653,6 +653,20 @@ static void quirk_no_aersid(struct pci_dev *pdev) DECLARE_PCI_FIXUP_CLASS_EARLY(PCI_VENDOR_ID_INTEL,PCI_ANY_ID ,PCI_CLASS_BRIDGE_PCI ,8 , quirk_no_aersid); +#if IS_ENABLED (CONFIG_VMD ) +/* +* VMD does not support legacy interrupts for downstream devices. +* So PCI_INTERRPUT_LINE needs to be initialized to 0 to ensure Quoted 2 lines
OS
+ * doesn't try to configure a legacy irq.s/legacy interrupts/INTx/
s/PCI_INTERRPUT_LINE/PCI_INTERRUPT_LINE/Quoted 8 lines
+ */
arch/x86/pci/fixup.c +4
-0+static void quirk_vmd_interrupt_line (struct pci_dev *dev) +{ +if (is_vmd (dev->bus)) + pci_write_config_byte(dev, PCI_INTERRUPT_LINE,0); +}
+DECLARE_PCI_FIXUP_HEADER(PCI_ANY_ID,PCI_ANY_ID,quirk_vmd_interrupt_line);A quirk for every PCI device, even on systems without VMD, seems like kind of a clumsy way to deal with this.
Quoted 9 lines
Conceptually, I would expect a host bridge driver (VMD acts like a host bridge in this case) to know whether it supports INTx, and if the driver knows it doesn't support INTx or it has no
_PRTor DT description of INTx routing to use, an attempt to configure INTx should just fail naturally.I don't claim this is how host bridge drivers actually work; I just think it's the way they *should* work.
Quoted 7 lines
Absolutely! This patch is fixing the issue in a wrong place. There are existing DT based host bridge drivers that disable INTx due to lack of hardware capability. The driver just need to nullify
pci_host_bridge::map_irqcallback.- Mani
For VMD as a host bridge, pci_host_bridge::map_irq is null. Even all VMD rootports' PCI_INTERRUPT_LINE registers are set to 0. Since VMD doesn't explicitly set PCI_INTERRUPT_LINE register to 0 for all of its sub-devices (i.e. NVMe), if some NVMes has non-zero value set for PCI_INTERRUPT_LINE (i.e. 0xff) then some software like SPDK can read it and make wrong assumption about INTx support.
Based Bjorn's and your suggestion, it might be better if VMD sets PCI_INTERRUPT_LINE register for all of its sub-devices during VMD enumeration.
-nirmal.
Quoted 66 lines · Nirmal Patel
Quoted 63 lines · On Thu, 25 Jul 2024 09:40:13 +0530 Manivannan Sadhasivam…
Quoted 47 lines · Bjorn Helgaas
Quoted 4 lines · Nirmal Patel
VMD does not support legacy interrupts for devices downstream from a VMD endpoint. So initialize the
PCI_INTERRUPT_LINEto 0 for these devices to ensure we don't try to set up a legacy irq for them.s/legacy interrupts/INTx/
s/legacy irq/INTx/Quoted 20 lines
Note: This patch was proposed by Jim, I am trying to upstream it.
Signed-off-by Jim Harris <[email protected]>Signed-off-by Nirmal Patel <[email protected]>--- arch/x86/pci/fixup.c | 14 ++++++++++++++ 1 file changed, 14 insertions(+)arch/x86/pci/fixup.c +4
-0@@ -653,6 +653,20 @@ static void quirk_no_aersid(struct pci_dev *pdev) DECLARE_PCI_FIXUP_CLASS_EARLY(PCI_VENDOR_ID_INTEL,PCI_ANY_ID ,PCI_CLASS_BRIDGE_PCI ,8 , quirk_no_aersid); +#if IS_ENABLED (CONFIG_VMD ) +/* +* VMD does not support legacy interrupts for downstream devices. +* So PCI_INTERRPUT_LINE needs to be initialized to 0 to ensure Quoted 2 lines
OS
+ * doesn't try to configure a legacy irq.s/legacy interrupts/INTx/
s/PCI_INTERRPUT_LINE/PCI_INTERRUPT_LINE/Quoted 8 lines
+ */
arch/x86/pci/fixup.c +4
-0+static void quirk_vmd_interrupt_line (struct pci_dev *dev) +{ +if (is_vmd (dev->bus)) + pci_write_config_byte(dev, PCI_INTERRUPT_LINE,0); +}
+DECLARE_PCI_FIXUP_HEADER(PCI_ANY_ID,PCI_ANY_ID,quirk_vmd_interrupt_line);A quirk for every PCI device, even on systems without VMD, seems like kind of a clumsy way to deal with this.
Quoted 9 lines
Conceptually, I would expect a host bridge driver (VMD acts like a host bridge in this case) to know whether it supports INTx, and if the driver knows it doesn't support INTx or it has no
_PRTor DT description of INTx routing to use, an attempt to configure INTx should just fail naturally.I don't claim this is how host bridge drivers actually work; I just think it's the way they *should* work.
Absolutely! This patch is fixing the issue in a wrong place. There are existing DT based host bridge drivers that disable INTx due to lack of hardware capability. The driver just need to nullify
pci_host_bridge::map_irqcallback.
Quoted 4 lines
Quoted 2 lines
- Mani
For VMD as a host bridge,
pci_host_bridge::map_irqis null. Even all VMD rootports'PCI_INTERRUPT_LINEregisters are set to 0.
If map_irq is already NULL, then how INTx is being configured? In your patch description:
"So initialize the PCI_INTERRUPT_LINE to 0 for these devices to ensure we don't try to set up a legacy irq for them."
Who is 'we'? For sure the PCI core wouldn't set INTx in your case. Does 'we' refer to device firmware?
Quoted 6 lines
Since VMD
doesn't explicitly setPCI_INTERRUPT_LINEregister to 0 for all of its sub-devices (i.e. NVMe), if some NVMes has non-zero value set forPCI_INTERRUPT_LINE(i.e.0xff) then some software like SPDK can read it and make wrong assumption about INTx support.
Is this statement is true (I haven't heard of before), then don't we need to set PCI_INTERRUPT_LINE to 0 for all devices irrespective of host bridge?
Quoted 4 lines
Based Bjorn's and your suggestion, it might be better if VMD sets
PCI_INTERRUPT_LINEregister for all of its sub-devices during VMD enumeration.
What about hotplug devices?
- Mani
Signature · 1 line
மணிவண்ணன் சதாசிவம்
Quoted 70 lines · On Tue, 30 Jul 2024 10:58:30 +0530 Manivannan Sadhasivam…
Quoted 69 lines · Nirmal Patel
Quoted 66 lines · On Thu, 25 Jul 2024 09:40:13 +0530 Manivannan Sadhasivam…
Quoted 50 lines · Bjorn Helgaas
Quoted 4 lines · Nirmal Patel
VMD does not support legacy interrupts for devices downstream from a VMD endpoint. So initialize the
PCI_INTERRUPT_LINEto 0 for these devices to ensure we don't try to set up a legacy irq for them.s/legacy interrupts/INTx/
s/legacy irq/INTx/Quoted 20 lines
Note: This patch was proposed by Jim, I am trying to upstream it.
Signed-off-by Jim Harris <[email protected]>Signed-off-by Nirmal Patel <[email protected]>--- arch/x86/pci/fixup.c | 14 ++++++++++++++ 1 file changed, 14 insertions(+)arch/x86/pci/fixup.c +3
-0@@ -653,6 +653,20 @@ static void quirk_no_aersid(struct pci_dev *pdev) DECLARE_PCI_FIXUP_CLASS_EARLY(PCI_VENDOR_ID_INTEL, PCI_ANY_ID, PCI_CLASS_BRIDGE_PCI, 8, quirk_no_aersid); +#if IS_ENABLED (CONFIG_VMD ) +/* +* VMD does not support legacy interrupts for downstream Quoted 4 lines
devices.
+ * SoPCI_INTERRPUT_LINEneeds to be initialized to 0 to ensure OS
+ * doesn't try to configure a legacy irq.s/legacy interrupts/INTx/
s/PCI_INTERRPUT_LINE/PCI_INTERRUPT_LINE/Quoted 8 lines
+ */
arch/x86/pci/fixup.c +4
-0+static void quirk_vmd_interrupt_line (struct pci_dev *dev) +{ +if (is_vmd (dev->bus)) + pci_write_config_byte(dev,
PCI_INTERRUPT_LINE, 0); +}
+DECLARE_PCI_FIXUP_HEADER(PCI_ANY_ID,PCI_ANY_ID,quirk_vmd_interrupt_line);A quirk for every PCI device, even on systems without VMD, seems like kind of a clumsy way to deal with this.
Quoted 9 lines
Conceptually, I would expect a host bridge driver (VMD acts like a host bridge in this case) to know whether it supports INTx, and if the driver knows it doesn't support INTx or it has no
_PRTor DT description of INTx routing to use, an attempt to configure INTx should just fail naturally.I don't claim this is how host bridge drivers actually work; I just think it's the way they *should* work.
Absolutely! This patch is fixing the issue in a wrong place. There are existing DT based host bridge drivers that disable INTx due to lack of hardware capability. The driver just need to nullify
pci_host_bridge::map_irqcallback.
Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
- Mani
For VMD as a host bridge,
pci_host_bridge::map_irqis null. Even all VMD rootports'PCI_INTERRUPT_LINEregisters are set to 0.If
map_irqis already NULL, then how INTx is being configured? In your patch description:
VMD uses MSIx.
Quoted 14 lines
"So initialize the
PCI_INTERRUPT_LINEto 0 for these devices to ensure we don't try to set up a legacy irq for them."Who is 'we'? For sure the PCI core wouldn't set INTx in your case. Does 'we' refer to device firmware?
Quoted 6 lines
Since VMD
doesn't explicitly setPCI_INTERRUPT_LINEregister to 0 for all of its sub-devices (i.e. NVMe), if some NVMes has non-zero value set forPCI_INTERRUPT_LINE(i.e.0xff) then some software like SPDK can read it and make wrong assumption about INTx support.
Quoted 3 lines
Is this statement is true (I haven't heard of before), then don't we need to set
PCI_INTERRUPT_LINEto 0 for all devices irrespective of host bridge?
Since VMD doesn't support legacy interrupt, BIOS sets PCI_INTERRUPT_LINE registers to 0 for all of the VMD rootports but not the NVMes'.
According to PCIe base specs, "Values in this register are programmed by system software and are system architecture specific. The Function itself does not use this value; rather the value in this register is used by device drivers and operating systems."
We had an issue raised on us sometime back because some SSDs have 0xff (i.e. Samsung) set to these registers by firmware and SPDK was reading them when SSDs were behind VMD which led them to believe VMD had INTx support enabled. After some testing, it made more sense to clear these registers for all of the VMD owned devices.
Quoted 7 lines
Quoted 5 lines
Based Bjorn's and your suggestion, it might be better if VMD sets
PCI_INTERRUPT_LINEregister for all of its sub-devices during VMD enumeration.What about hotplug devices?
That is a good question and because of that I thought of putting the fix in fixup.c. But I am open to your suggestion since fixup is not the right place.
Thanks
- nirmal
Quoted 3 lines
- Mani
Quoted 107 lines · Nirmal Patel
Quoted 70 lines · On Tue, 30 Jul 2024 10:58:30 +0530 Manivannan Sadhasivam…
Quoted 69 lines · Nirmal Patel
Quoted 66 lines · On Thu, 25 Jul 2024 09:40:13 +0530 Manivannan Sadhasivam…
Quoted 50 lines · Bjorn Helgaas
Quoted 4 lines · Nirmal Patel
VMD does not support legacy interrupts for devices downstream from a VMD endpoint. So initialize the
PCI_INTERRUPT_LINEto 0 for these devices to ensure we don't try to set up a legacy irq for them.s/legacy interrupts/INTx/
s/legacy irq/INTx/Quoted 20 lines
Note: This patch was proposed by Jim, I am trying to upstream it.
Signed-off-by Jim Harris <[email protected]>Signed-off-by Nirmal Patel <[email protected]>--- arch/x86/pci/fixup.c | 14 ++++++++++++++ 1 file changed, 14 insertions(+)arch/x86/pci/fixup.c +3
-0@@ -653,6 +653,20 @@ static void quirk_no_aersid(struct pci_dev *pdev) DECLARE_PCI_FIXUP_CLASS_EARLY(PCI_VENDOR_ID_INTEL, PCI_ANY_ID, PCI_CLASS_BRIDGE_PCI, 8, quirk_no_aersid); +#if IS_ENABLED (CONFIG_VMD ) +/* +* VMD does not support legacy interrupts for downstream Quoted 4 lines
devices.
+ * SoPCI_INTERRPUT_LINEneeds to be initialized to 0 to ensure OS
+ * doesn't try to configure a legacy irq.s/legacy interrupts/INTx/
s/PCI_INTERRPUT_LINE/PCI_INTERRUPT_LINE/Quoted 8 lines
+ */
arch/x86/pci/fixup.c +4
-0+static void quirk_vmd_interrupt_line (struct pci_dev *dev) +{ +if (is_vmd (dev->bus)) + pci_write_config_byte(dev,
PCI_INTERRUPT_LINE, 0); +}
+DECLARE_PCI_FIXUP_HEADER(PCI_ANY_ID,PCI_ANY_ID,quirk_vmd_interrupt_line);A quirk for every PCI device, even on systems without VMD, seems like kind of a clumsy way to deal with this.
Quoted 9 lines
Conceptually, I would expect a host bridge driver (VMD acts like a host bridge in this case) to know whether it supports INTx, and if the driver knows it doesn't support INTx or it has no
_PRTor DT description of INTx routing to use, an attempt to configure INTx should just fail naturally.I don't claim this is how host bridge drivers actually work; I just think it's the way they *should* work.
Absolutely! This patch is fixing the issue in a wrong place. There are existing DT based host bridge drivers that disable INTx due to lack of hardware capability. The driver just need to nullify
pci_host_bridge::map_irqcallback.Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
- Mani
For VMD as a host bridge,
pci_host_bridge::map_irqis null. Even all VMD rootports'PCI_INTERRUPT_LINEregisters are set to 0.If
map_irqis already NULL, then how INTx is being configured? In your patch description:VMD uses MSIx.
Quoted 14 lines
"So initialize the
PCI_INTERRUPT_LINEto 0 for these devices to ensure we don't try to set up a legacy irq for them."Who is 'we'? For sure the PCI core wouldn't set INTx in your case. Does 'we' refer to device firmware?
Quoted 6 lines
Since VMD
doesn't explicitly setPCI_INTERRUPT_LINEregister to 0 for all of its sub-devices (i.e. NVMe), if some NVMes has non-zero value set forPCI_INTERRUPT_LINE(i.e.0xff) then some software like SPDK can read it and make wrong assumption about INTx support.Quoted 3 lines
Is this statement is true (I haven't heard of before), then don't we need to set
PCI_INTERRUPT_LINEto 0 for all devices irrespective of host bridge?Since VMD doesn't support legacy interrupt, BIOS sets
PCI_INTERRUPT_LINEregisters to 0 for all of the VMD rootports but not the NVMes'.According to PCIe base specs, "Values in this register are programmed by system software and are system architecture specific. The Function itself does not use this value; rather the value in this register is used by device drivers and operating systems."
Quoted 6 lines
We had an issue raised on us sometime back because some SSDs have
0xff(i.e. Samsung) set to these registers by firmware and SPDK was reading them when SSDs were behind VMD which led them to believe VMD had INTx support enabled. After some testing, it made more sense to clear these registers for all of the VMD owned devices.
This is a valuable information that should've been present in the patch description. Now I can understand the intention of your patch. Previously I couldn't.
Quoted 11 lines
Quoted 7 lines
Quoted 5 lines
Based Bjorn's and your suggestion, it might be better if VMD sets
PCI_INTERRUPT_LINEregister for all of its sub-devices during VMD enumeration.What about hotplug devices?
That is a good question and because of that I thought of putting the fix in
fixup.c. But I am open to your suggestion since fixup is not the right place.
How about the below change?
drivers/pci/irq.c +7 -0
@@ -147,6 +147,13 @@ void pci_assign_irq(struct pci_dev *dev)struct pci_host_bridge *hbrg =pci_find_host_bridge (dev->bus);if (!(hbrg->map_irq)) { +/* +* Some userspace applications like SPDK reads +* PCI_INTERRUPT_LINE to decide whether INTx is enabled or not. +* So write 0 to make sure they understand that INTx is disabled +* for the device. +*/ +pci_write_config_byte (dev,PCI_INTERRUPT_LINE ,0 );pci_dbg (dev,"runtime IRQ mapping not provided by arch\n" );return ; }
So this sets PCI_INTERRUPT_LINE to 0 for _all_ devices that don't support INTx. As per your explanation above, the issue you are seeing is not just applicable to VMD, but for all devices.
- Mani
Signature · 1 line
மணிவண்ணன் சதாசிவம்
Quoted 157 lines · On Wed, 31 Jul 2024 08:37:39 +0530 Manivannan Sadhasivam…
Quoted 109 lines · Nirmal Patel
Quoted 71 lines · On Tue, 30 Jul 2024 10:58:30 +0530 Manivannan Sadhasivam…
Quoted 70 lines · Nirmal Patel
Quoted 67 lines · On Thu, 25 Jul 2024 09:40:13 +0530 Manivannan Sadhasivam…
Quoted 50 lines · Bjorn Helgaas
Quoted 4 lines · Nirmal Patel
VMD does not support legacy interrupts for devices downstream from a VMD endpoint. So initialize the
PCI_INTERRUPT_LINEto 0 for these devices to ensure we don't try to set up a legacy irq for them.s/legacy interrupts/INTx/
s/legacy irq/INTx/Quoted 20 lines
Note: This patch was proposed by Jim, I am trying to upstream it.
Signed-off-by Jim Harris <[email protected]>Signed-off-by Nirmal Patel <[email protected]>--- arch/x86/pci/fixup.c | 14 ++++++++++++++ 1 file changed, 14 insertions(+)arch/x86/pci/fixup.c +3
-0@@ -653,6 +653,20 @@ static void quirk_no_aersid(struct pci_dev *pdev) DECLARE_PCI_FIXUP_CLASS_EARLY(PCI_VENDOR_ID_INTEL, PCI_ANY_ID, PCI_CLASS_BRIDGE_PCI, 8, quirk_no_aersid); +#if IS_ENABLED (CONFIG_VMD ) +/* +* VMD does not support legacy interrupts for downstream Quoted 4 lines
devices.
+ * SoPCI_INTERRPUT_LINEneeds to be initialized to 0 to ensure OS
+ * doesn't try to configure a legacy irq.s/legacy interrupts/INTx/
s/PCI_INTERRPUT_LINE/PCI_INTERRUPT_LINE/Quoted 8 lines
+ */
arch/x86/pci/fixup.c +4
-0+static void quirk_vmd_interrupt_line (struct pci_dev *dev) +{ +if (is_vmd (dev->bus)) + pci_write_config_byte(dev,
PCI_INTERRUPT_LINE, 0); +} +DECLARE_PCI_FIXUP_HEADER(PCI_ANY_ID,PCI_ANY_ID,quirk_vmd_interrupt_line);A quirk for every PCI device, even on systems without VMD, seems like kind of a clumsy way to deal with this.
Quoted 9 lines
Conceptually, I would expect a host bridge driver (VMD acts like a host bridge in this case) to know whether it supports INTx, and if the driver knows it doesn't support INTx or it has no
_PRTor DT description of INTx routing to use, an attempt to configure INTx should just fail naturally.I don't claim this is how host bridge drivers actually work; I just think it's the way they *should* work.
Absolutely! This patch is fixing the issue in a wrong place. There are existing DT based host bridge drivers that disable INTx due to lack of hardware capability. The driver just need to nullify
pci_host_bridge::map_irqcallback.Quoted 8 lines
Quoted 5 lines
Quoted 2 lines
- Mani
For VMD as a host bridge,
pci_host_bridge::map_irqis null. Even all VMD rootports'PCI_INTERRUPT_LINEregisters are set to 0.If
map_irqis already NULL, then how INTx is being configured? In your patch description:VMD uses MSIx.
Quoted 14 lines
"So initialize the
PCI_INTERRUPT_LINEto 0 for these devices to ensure we don't try to set up a legacy irq for them."Who is 'we'? For sure the PCI core wouldn't set INTx in your case. Does 'we' refer to device firmware?
Quoted 6 lines
Since VMD
doesn't explicitly setPCI_INTERRUPT_LINEregister to 0 for all of its sub-devices (i.e. NVMe), if some NVMes has non-zero value set forPCI_INTERRUPT_LINE(i.e.0xff) then some software like SPDK can read it and make wrong assumption about INTx support.Quoted 3 lines
Is this statement is true (I haven't heard of before), then don't we need to set
PCI_INTERRUPT_LINEto 0 for all devices irrespective of host bridge?Since VMD doesn't support legacy interrupt, BIOS sets
PCI_INTERRUPT_LINEregisters to 0 for all of the VMD rootports but not the NVMes'.According to PCIe base specs, "Values in this register are programmed by system software and are system architecture specific. The Function itself does not use this value; rather the value in this register is used by device drivers and operating systems."
Quoted 6 lines
We had an issue raised on us sometime back because some SSDs have
0xff(i.e. Samsung) set to these registers by firmware and SPDK was reading them when SSDs were behind VMD which led them to believe VMD had INTx support enabled. After some testing, it made more sense to clear these registers for all of the VMD owned devices.This is a valuable information that should've been present in the patch description. Now I can understand the intention of your patch. Previously I couldn't.
Quoted 11 lines
Quoted 7 lines
Quoted 5 lines
Based Bjorn's and your suggestion, it might be better if VMD sets
PCI_INTERRUPT_LINEregister for all of its sub-devices during VMD enumeration.What about hotplug devices?
That is a good question and because of that I thought of putting the fix in
fixup.c. But I am open to your suggestion since fixup is not the right place.How about the below change?
drivers/pci/irq.c +3
-0@@ -147,6 +147,13 @@ void pci_assign_irq(struct pci_dev *dev)struct pci_host_bridge *hbrg =pci_find_host_bridge (dev->bus);if (!(hbrg->map_irq)) { +/* +* Some userspace applications like SPDK reads +* PCI_INTERRUPT_LINE to decide whether INTx is enabled or not.
+ * So write 0 to make sure they understand that INTx
is disableddrivers/pci/irq.c +3
-0+* for the device. +*/ +pci_write_config_byte (dev,PCI_INTERRUPT_LINE ,0 ); pci_dbg(dev, "runtime IRQ mapping not provided byarch\n"); return; }
Quoted 6 lines
So this sets
PCI_INTERRUPT_LINEto 0 for _all_ devices that don't support INTx. As per your explanation above, the issue you are seeing is not just applicable to VMD, but for all devices.- Mani
Thanks for the suggestion. Let me test the changes.
-nirmal
Quoted 45 lines · On Wed, 24 Jul 2024 14:10:30 -0500 Bjorn Helgaas <[email protected]> wrote:
Quoted 3 lines · Nirmal Patel
VMD does not support legacy interrupts for devices downstream from a VMD endpoint. So initialize the
PCI_INTERRUPT_LINEto 0 for these devices to ensure we don't try to set up a legacy irq for them.s/legacy interrupts/INTx/
s/legacy irq/INTx/Quoted 13 lines
Note: This patch was proposed by Jim, I am trying to upstream it.
Signed-off-by Jim Harris <[email protected]>Signed-off-by Nirmal Patel <[email protected]>--- arch/x86/pci/fixup.c | 14 ++++++++++++++ 1 file changed, 14 insertions(+)arch/x86/pci/fixup.c +0
-0arch/x86/pci/fixup.c static void quirk_no_aersid(struct pci_dev · line 660
Quoted 8 lines
arch/x86/pci/fixup.c +5
-0@@ -653,6 +653,20 @@ static void quirk_no_aersid(struct pci_dev *pdev) DECLARE_PCI_FIXUP_CLASS_EARLY(PCI_VENDOR_ID_INTEL,PCI_ANY_ID ,PCI_CLASS_BRIDGE_PCI ,8 , quirk_no_aersid); +#if IS_ENABLED (CONFIG_VMD ) +/* +* VMD does not support legacy interrupts for downstream devices. +* So PCI_INTERRPUT_LINE needs to be initialized to 0 to ensure OS +* doesn't try to configure a legacy irq. s/legacy interrupts/INTx/
s/PCI_INTERRPUT_LINE/PCI_INTERRUPT_LINE/arch/x86/pci/fixup.c static void quirk_no_aersid(struct pci_dev *pdev)
Quoted 8 lines
arch/x86/pci/fixup.c +7
-0+*/ +static void quirk_vmd_interrupt_line (struct pci_dev *dev) +{ +if (is_vmd (dev->bus)) +pci_write_config_byte (dev,PCI_INTERRUPT_LINE ,0 ); +} +DECLARE_PCI_FIXUP_HEADER (PCI_ANY_ID ,PCI_ANY_ID , quirk_vmd_interrupt_line);A quirk for every PCI device, even on systems without VMD, seems like kind of a clumsy way to deal with this.
Quoted 8 lines
Conceptually, I would expect a host bridge driver (VMD acts like a host bridge in this case) to know whether it supports INTx, and if the driver knows it doesn't support INTx or it has no
_PRTor DT description of INTx routing to use, an attempt to configure INTx should just fail naturally.I don't claim this is how host bridge drivers actually work; I just think it's the way they *should* work.
For VMD as a host bridge, pci_host_bridge::map_irq is null. Even all VMD rootports' PCI_INTERRUPT_LINE registers are set to 0. Since VMD doesn't explicitly set PCI_INTERRUPT_LINE register to 0 for all of its sub-devices (i.e. NVMe), if some NVMes has non-zero value set for PCI_INTERRUPT_LINE (i.e. 0xff) then some software like SPDK can read it and make wrong assumption about INTx support.
Based your suggestion, it might be better if VMD sets PCI_INTERRUPT_LINE register to 0 for all of its sub-devices during VMD enumeration.
-nirmal.
Quoted 9 lines
Quoted 9 lines
arch/x86/pci/fixup.c +2
-0+#endif +static void quirk_intel_th_dnv (struct pci_dev *dev) {struct resource *r = &dev->resource[4];
Signature · 1 line
2.39.1