Skip to contentSkip to conversation

PCI/MSI: Set device flag indicating only 32-bit MSI support

PATCH Superseded Vidya Sagar 2 participants 9 messages Nov 17, 2020 – Dec 4, 2020 PCI

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 PCI/MSI: Set device flag indicating only 32-bit MSI support

There are devices (Ex:- Marvell SATA controller) that don't support 64-bit MSIs and the same is advertised through their MSI capability register. Set no_64bit_msi flag explicitly for such devices in the MSI setup code so that the msi_verify_entries() API would catch if the MSI arch code tries to use 64-bit MSI.

---
 drivers/pci/msi.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
drivers/pci/msi.c +4 -2
index d52d118979a6..af49da28854e 100644
@@ -581,10 +581,12 @@ msi_setup_entry(struct pci_dev *dev, int nvec, struct irq_affinity *affd)
 	entry->msi_attrib.multi_cap	= (control & PCI_MSI_FLAGS_QMASK) >> 1;
 	entry->msi_attrib.multiple	= ilog2(__roundup_pow_of_two(nvec));
 
-	if (control & PCI_MSI_FLAGS_64BIT)
+	if (control & PCI_MSI_FLAGS_64BIT) {
 		entry->mask_pos = dev->msi_cap + PCI_MSI_MASK_64;
-	else
+	} else {
 		entry->mask_pos = dev->msi_cap + PCI_MSI_MASK_32;
+		dev->no_64bit_msi = 1;
+	}
 
 	/* Save the initial mask status */
 	if (entry->msi_attrib.maskbit)
Signature · 1 line
2.17.1
#1 lore raw
Quoted 3 lines · Vidya Sagar

There are devices (Ex:- Marvell SATA controller) that don't support 64-bit MSIs and the same is advertised through their MSI capability register.

I *think* you're saying these devices behave correctly per spec: they don't support 64-bit MSI, and they don't advertise support for 64-bit MSI. Right?

Quoted 3 lines

Set no_64bit_msi flag explicitly for such devices in the MSI setup code so that the msi_verify_entries() API would catch if the MSI arch code tries to use 64-bit MSI.

And you want msi_verify_entries() to catch attempts by the arch code to assign a 64-bit MSI address?

That sounds OK, but the error message ("Device has broken 64-bit MSI") is not appropriate if the device is actually *not* broken.

Quoted 27 lines
---
 drivers/pci/msi.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
drivers/pci/msi.c +4 -2
index d52d118979a6..af49da28854e 100644
@@ -581,10 +581,12 @@ msi_setup_entry(struct pci_dev *dev, int nvec, struct irq_affinity *affd)
 	entry->msi_attrib.multi_cap	= (control & PCI_MSI_FLAGS_QMASK) >> 1;
 	entry->msi_attrib.multiple	= ilog2(__roundup_pow_of_two(nvec));
 
-	if (control & PCI_MSI_FLAGS_64BIT)
+	if (control & PCI_MSI_FLAGS_64BIT) {
 		entry->mask_pos = dev->msi_cap + PCI_MSI_MASK_64;
-	else
+	} else {
 		entry->mask_pos = dev->msi_cap + PCI_MSI_MASK_32;
+		dev->no_64bit_msi = 1;
+	}
 
 	/* Save the initial mask status */
 	if (entry->msi_attrib.maskbit)
Signature · 1 line
2.17.1
#2 lore raw
Quoted 11 lines · Bjorn Helgaas

External email: Use caution opening links or attachments

Quoted 3 lines · Vidya Sagar

There are devices (Ex:- Marvell SATA controller) that don't support 64-bit MSIs and the same is advertised through their MSI capability register.

I *think* you're saying these devices behave correctly per spec: they don't support 64-bit MSI, and they don't advertise support for 64-bit MSI. Right?

Yes. That is what I intended to say.

Quoted 7 lines
Quoted 4 lines

Set no_64bit_msi flag explicitly for such devices in the MSI setup code so that the msi_verify_entries() API would catch if the MSI arch code tries to use 64-bit MSI.

And you want msi_verify_entries() to catch attempts by the arch code to assign a 64-bit MSI address?

Yes.

Quoted 3 lines

That sounds OK, but the error message ("Device has broken 64-bit MSI") is not appropriate if the device is actually *not* broken.

Ok. I didn't change the existing error message. I'll change it to cover both the scenarios i.e. either the device is broken or the device doesn't really support 64-bit MSI.

Quoted 28 lines
Quoted 28 lines
---
  drivers/pci/msi.c | 6 ++++--
  1 file changed, 4 insertions(+), 2 deletions(-)
drivers/pci/msi.c +4 -3
index d52d118979a6..af49da28854e 100644
@@ -581,10 +581,12 @@ msi_setup_entry(struct pci_dev *dev, int nvec, struct irq_affinity *affd)
       entry->msi_attrib.multi_cap     = (control & PCI_MSI_FLAGS_QMASK) >> 1;
       entry->msi_attrib.multiple      = ilog2(__roundup_pow_of_two(nvec));
 
-     if (control & PCI_MSI_FLAGS_64BIT)
+     if (control & PCI_MSI_FLAGS_64BIT) {
               entry->mask_pos = dev->msi_cap + PCI_MSI_MASK_64;
-     else
+     } else {
               entry->mask_pos = dev->msi_cap + PCI_MSI_MASK_32;
+             dev->no_64bit_msi = 1;
+     }
 
       /* Save the initial mask status */
       if (entry->msi_attrib.maskbit)
--

2.17.1

#3 lore raw

PATCH v2 1/1 PCI/MSI: Set device flag indicating only 32-bit MSI support

In reply to Vidya Sagar

There are devices (Ex:- Marvell SATA controller) that don't support 64-bit MSIs and the same is advertised through their MSI capability register. Set no_64bit_msi flag explicitly for such devices in the MSI setup code so that the msi_verify_entries() API would catch if the MSI arch code tries to use 64-bit MSI.

---
V2:
* Addressed Bjorn's comment and changed the error message

 drivers/pci/msi.c | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)
drivers/pci/msi.c +7 -4
index d52d118979a6..8de5ba6b4a59 100644
@@ -581,10 +581,12 @@ msi_setup_entry(struct pci_dev *dev, int nvec, struct irq_affinity *affd)
 	entry->msi_attrib.multi_cap	= (control & PCI_MSI_FLAGS_QMASK) >> 1;
 	entry->msi_attrib.multiple	= ilog2(__roundup_pow_of_two(nvec));
 
-	if (control & PCI_MSI_FLAGS_64BIT)
+	if (control & PCI_MSI_FLAGS_64BIT) {
 		entry->mask_pos = dev->msi_cap + PCI_MSI_MASK_64;
-	else
+	} else {
 		entry->mask_pos = dev->msi_cap + PCI_MSI_MASK_32;
+		dev->no_64bit_msi = 1;
+	}
 
 	/* Save the initial mask status */
 	if (entry->msi_attrib.maskbit)
@@ -602,8 +604,9 @@ static int msi_verify_entries(struct pci_dev *dev)
 	for_each_pci_msi_entry(entry, dev) {
 		if (!dev->no_64bit_msi || !entry->msg.address_hi)
 			continue;
-		pci_err(dev, "Device has broken 64-bit MSI but arch"
-			" tried to assign one above 4G\n");
+		pci_err(dev, "Device has either broken 64-bit MSI or "
+			"only 32-bit MSI support but "
+			"arch tried to assign one above 4G\n");
 		return -EIO;
 	}
 	return 0;
Signature · 1 line
2.17.1
#4 lore raw

Hi Bjorn,
Do you have any further comments for this patch?

Thanks,
Vidya Sagar

Quoted 46 lines · Vidya Sagar

There are devices (Ex:- Marvell SATA controller) that don't support 64-bit MSIs and the same is advertised through their MSI capability register. Set no_64bit_msi flag explicitly for such devices in the MSI setup code so that the msi_verify_entries() API would catch if the MSI arch code tries to use 64-bit MSI.

---
V2:
* Addressed Bjorn's comment and changed the error message

  drivers/pci/msi.c | 11 +++++++----
  1 file changed, 7 insertions(+), 4 deletions(-)
drivers/pci/msi.c +7 -4
index d52d118979a6..8de5ba6b4a59 100644
@@ -581,10 +581,12 @@ msi_setup_entry(struct pci_dev *dev, int nvec, struct irq_affinity *affd)
  	entry->msi_attrib.multi_cap	= (control & PCI_MSI_FLAGS_QMASK) >> 1;
  	entry->msi_attrib.multiple	= ilog2(__roundup_pow_of_two(nvec));
 
-	if (control & PCI_MSI_FLAGS_64BIT)
+	if (control & PCI_MSI_FLAGS_64BIT) {
  		entry->mask_pos = dev->msi_cap + PCI_MSI_MASK_64;
-	else
+	} else {
  		entry->mask_pos = dev->msi_cap + PCI_MSI_MASK_32;
+		dev->no_64bit_msi = 1;
+	}
 
  	/* Save the initial mask status */
  	if (entry->msi_attrib.maskbit)
@@ -602,8 +604,9 @@ static int msi_verify_entries(struct pci_dev *dev)
  	for_each_pci_msi_entry(entry, dev) {
  		if (!dev->no_64bit_msi || !entry->msg.address_hi)
  			continue;
-		pci_err(dev, "Device has broken 64-bit MSI but arch"
-			" tried to assign one above 4G\n");
+		pci_err(dev, "Device has either broken 64-bit MSI or "
+			"only 32-bit MSI support but "
+			"arch tried to assign one above 4G\n");
  		return -EIO;
  	}
  	return 0;
#5 lore raw
In reply to Vidya Sagar
Quoted 5 lines · Vidya Sagar

There are devices (Ex:- Marvell SATA controller) that don't support 64-bit MSIs and the same is advertised through their MSI capability register. Set no_64bit_msi flag explicitly for such devices in the MSI setup code so that the msi_verify_entries() API would catch if the MSI arch code tries to use 64-bit MSI.

This seems good to me. I'll post a possible revision to set dev->no_64bit_msi in the device enumeration path instead of in the IRQ allocation path, since it's really a property of the device, not of the msi_desc.

I like the extra checking this gives us. Was this prompted by tripping over something, or is it something you noticed by code reading? If the former, a hint about what was wrong and how it's being fixed would be useful.

Quoted 42 lines

---
V2:
* Addressed Bjorn's comment and changed the error message

 drivers/pci/msi.c | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)
drivers/pci/msi.c +7 -4
index d52d118979a6..8de5ba6b4a59 100644
@@ -581,10 +581,12 @@ msi_setup_entry(struct pci_dev *dev, int nvec, struct irq_affinity *affd)
 	entry->msi_attrib.multi_cap	= (control & PCI_MSI_FLAGS_QMASK) >> 1;
 	entry->msi_attrib.multiple	= ilog2(__roundup_pow_of_two(nvec));
 
-	if (control & PCI_MSI_FLAGS_64BIT)
+	if (control & PCI_MSI_FLAGS_64BIT) {
 		entry->mask_pos = dev->msi_cap + PCI_MSI_MASK_64;
-	else
+	} else {
 		entry->mask_pos = dev->msi_cap + PCI_MSI_MASK_32;
+		dev->no_64bit_msi = 1;
+	}
 
 	/* Save the initial mask status */
 	if (entry->msi_attrib.maskbit)
@@ -602,8 +604,9 @@ static int msi_verify_entries(struct pci_dev *dev)
 	for_each_pci_msi_entry(entry, dev) {
 		if (!dev->no_64bit_msi || !entry->msg.address_hi)
 			continue;
-		pci_err(dev, "Device has broken 64-bit MSI but arch"
-			" tried to assign one above 4G\n");
+		pci_err(dev, "Device has either broken 64-bit MSI or "
+			"only 32-bit MSI support but "
+			"arch tried to assign one above 4G\n");
 		return -EIO;
 	}
 	return 0;
Signature · 1 line
2.17.1
#6 lore raw
Quoted 15 lines · Bjorn Helgaas

External email: Use caution opening links or attachments

Quoted 5 lines · Vidya Sagar

There are devices (Ex:- Marvell SATA controller) that don't support 64-bit MSIs and the same is advertised through their MSI capability register. Set no_64bit_msi flag explicitly for such devices in the MSI setup code so that the msi_verify_entries() API would catch if the MSI arch code tries to use 64-bit MSI.

This seems good to me. I'll post a possible revision to set dev->no_64bit_msi in the device enumeration path instead of in the IRQ allocation path, since it's really a property of the device, not of the msi_desc.

Quoted 4 lines

I like the extra checking this gives us. Was this prompted by tripping over something, or is it something you noticed by code reading? If the former, a hint about what was wrong and how it's being fixed would be useful.

I observed functionality issue with Marvell SATA controller (1b4b:9171) when the allocated MSI target address was a 64-bit address. I mentioned the Marvell SATA controller as an example in the commit message.

Thanks,
Vidya Sagar

Quoted 43 lines
Quoted 43 lines

---
V2:
* Addressed Bjorn's comment and changed the error message

  drivers/pci/msi.c | 11 +++++++----
  1 file changed, 7 insertions(+), 4 deletions(-)
drivers/pci/msi.c +7 -5
index d52d118979a6..8de5ba6b4a59 100644
@@ -581,10 +581,12 @@ msi_setup_entry(struct pci_dev *dev, int nvec, struct irq_affinity *affd)
       entry->msi_attrib.multi_cap     = (control & PCI_MSI_FLAGS_QMASK) >> 1;
       entry->msi_attrib.multiple      = ilog2(__roundup_pow_of_two(nvec));
 
-     if (control & PCI_MSI_FLAGS_64BIT)
+     if (control & PCI_MSI_FLAGS_64BIT) {
               entry->mask_pos = dev->msi_cap + PCI_MSI_MASK_64;
-     else
+     } else {
               entry->mask_pos = dev->msi_cap + PCI_MSI_MASK_32;
+             dev->no_64bit_msi = 1;
+     }
 
       /* Save the initial mask status */
       if (entry->msi_attrib.maskbit)
@@ -602,8 +604,9 @@ static int msi_verify_entries(struct pci_dev *dev)
       for_each_pci_msi_entry(entry, dev) {
               if (!dev->no_64bit_msi || !entry->msg.address_hi)
                       continue;
-             pci_err(dev, "Device has broken 64-bit MSI but arch"
-                     " tried to assign one above 4G\n");
+             pci_err(dev, "Device has either broken 64-bit MSI or "
+                     "only 32-bit MSI support but "
+                     "arch tried to assign one above 4G\n");
               return -EIO;
       }
       return 0;
--

2.17.1

#7 lore raw
Quoted 13 lines · Vidya Sagar
Quoted 12 lines · Bjorn Helgaas
Quoted 5 lines · Vidya Sagar

There are devices (Ex:- Marvell SATA controller) that don't support 64-bit MSIs and the same is advertised through their MSI capability register. Set no_64bit_msi flag explicitly for such devices in the MSI setup code so that the msi_verify_entries() API would catch if the MSI arch code tries to use 64-bit MSI.

This seems good to me. I'll post a possible revision to set dev->no_64bit_msi in the device enumeration path instead of in the IRQ allocation path, since it's really a property of the device, not of the msi_desc.

Quoted 7 lines
Quoted 4 lines

I like the extra checking this gives us. Was this prompted by tripping over something, or is it something you noticed by code reading? If the former, a hint about what was wrong and how it's being fixed would be useful.

I observed functionality issue with Marvell SATA controller (1b4b:9171) when the allocated MSI target address was a 64-bit address. I mentioned the Marvell SATA controller as an example in the commit message.

I know you mentioned the Marvell controller, but apparently that device is working perfectly correctly: it does not support 64-bit MSI, and it does not advertise support for 64-bit MSI.

So if there's a functionality issue, that means something is wrong in Linux that caused us to assign a 64-bit MSI address to it. *That* issue is what I want to know about. Your patch only warns about the issue; it doesn't fix it.

I don't think there's any point in specifically mentioning the Marvell device if it is working correctly, because the same issue should affect *any* device that doesn't support 64-bit MSI. But if there's some arch code that incorrectly assigns a 64-bit address, it would definitely be useful to specify the arch. And hopefully there's a fix for that arch code, too.

Quoted 42 lines
Quoted 42 lines
Quoted 42 lines

---
V2:
* Addressed Bjorn's comment and changed the error message

  drivers/pci/msi.c | 11 +++++++----
  1 file changed, 7 insertions(+), 4 deletions(-)
drivers/pci/msi.c +7 -5
index d52d118979a6..8de5ba6b4a59 100644
@@ -581,10 +581,12 @@ msi_setup_entry(struct pci_dev *dev, int nvec, struct irq_affinity *affd)
       entry->msi_attrib.multi_cap     = (control & PCI_MSI_FLAGS_QMASK) >> 1;
       entry->msi_attrib.multiple      = ilog2(__roundup_pow_of_two(nvec));
 
-     if (control & PCI_MSI_FLAGS_64BIT)
+     if (control & PCI_MSI_FLAGS_64BIT) {
               entry->mask_pos = dev->msi_cap + PCI_MSI_MASK_64;
-     else
+     } else {
               entry->mask_pos = dev->msi_cap + PCI_MSI_MASK_32;
+             dev->no_64bit_msi = 1;
+     }
 
       /* Save the initial mask status */
       if (entry->msi_attrib.maskbit)
@@ -602,8 +604,9 @@ static int msi_verify_entries(struct pci_dev *dev)
       for_each_pci_msi_entry(entry, dev) {
               if (!dev->no_64bit_msi || !entry->msg.address_hi)
                       continue;
-             pci_err(dev, "Device has broken 64-bit MSI but arch"
-                     " tried to assign one above 4G\n");
+             pci_err(dev, "Device has either broken 64-bit MSI or "
+                     "only 32-bit MSI support but "
+                     "arch tried to assign one above 4G\n");
               return -EIO;
       }
       return 0;
--

2.17.1

#8 lore raw
Quoted 25 lines · Bjorn Helgaas

External email: Use caution opening links or attachments

Quoted 13 lines · Vidya Sagar
Quoted 12 lines · Bjorn Helgaas
Quoted 5 lines · Vidya Sagar

There are devices (Ex:- Marvell SATA controller) that don't support 64-bit MSIs and the same is advertised through their MSI capability register. Set no_64bit_msi flag explicitly for such devices in the MSI setup code so that the msi_verify_entries() API would catch if the MSI arch code tries to use 64-bit MSI.

This seems good to me. I'll post a possible revision to set dev->no_64bit_msi in the device enumeration path instead of in the IRQ allocation path, since it's really a property of the device, not of the msi_desc.

Quoted 7 lines
Quoted 4 lines

I like the extra checking this gives us. Was this prompted by tripping over something, or is it something you noticed by code reading? If the former, a hint about what was wrong and how it's being fixed would be useful.

I observed functionality issue with Marvell SATA controller (1b4b:9171) when the allocated MSI target address was a 64-bit address. I mentioned the Marvell SATA controller as an example in the commit message.

Quoted 8 lines

I know you mentioned the Marvell controller, but apparently that device is working perfectly correctly: it does not support 64-bit MSI, and it does not advertise support for 64-bit MSI.

So if there's a functionality issue, that means something is wrong in Linux that caused us to assign a 64-bit MSI address to it. *That* issue is what I want to know about. Your patch only warns about the issue; it doesn't fix it.

The issue is in the DWC code. I pushed a generic patch to fix that issue by specifying the limit of MSI target address to 32-bit region. Patch is up for review at http://patchwork.ozlabs.org/project/linux-pci/patch/[email protected]/

Quoted 50 lines

I don't think there's any point in specifically mentioning the Marvell device if it is working correctly, because the same issue should affect *any* device that doesn't support 64-bit MSI. But if there's some arch code that incorrectly assigns a 64-bit address, it would definitely be useful to specify the arch. And hopefully there's a fix for that arch code, too.

Quoted 42 lines
Quoted 42 lines
Quoted 42 lines

---
V2:
* Addressed Bjorn's comment and changed the error message

   drivers/pci/msi.c | 11 +++++++----
   1 file changed, 7 insertions(+), 4 deletions(-)
drivers/pci/msi.c +7 -5
index d52d118979a6..8de5ba6b4a59 100644
@@ -581,10 +581,12 @@ msi_setup_entry(struct pci_dev *dev, int nvec, struct irq_affinity *affd)
        entry->msi_attrib.multi_cap     = (control & PCI_MSI_FLAGS_QMASK) >> 1;
        entry->msi_attrib.multiple      = ilog2(__roundup_pow_of_two(nvec));
 
-     if (control & PCI_MSI_FLAGS_64BIT)
+     if (control & PCI_MSI_FLAGS_64BIT) {
                entry->mask_pos = dev->msi_cap + PCI_MSI_MASK_64;
-     else
+     } else {
                entry->mask_pos = dev->msi_cap + PCI_MSI_MASK_32;
+             dev->no_64bit_msi = 1;
+     }
 
        /* Save the initial mask status */
        if (entry->msi_attrib.maskbit)
@@ -602,8 +604,9 @@ static int msi_verify_entries(struct pci_dev *dev)
        for_each_pci_msi_entry(entry, dev) {
                if (!dev->no_64bit_msi || !entry->msg.address_hi)
                        continue;
-             pci_err(dev, "Device has broken 64-bit MSI but arch"
-                     " tried to assign one above 4G\n");
+             pci_err(dev, "Device has either broken 64-bit MSI or "
+                     "only 32-bit MSI support but "
+                     "arch tried to assign one above 4G\n");
                return -EIO;
        }
        return 0;
--

2.17.1