Skip to contentSkip to conversation

pci: do not try to assign irq 255

PATCH In review Hannes Reinecke 6 participants 17 messages Feb 18, 2013 – Sep 10, 2013

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 pci: do not try to assign irq 255

The PCI config space reseves a byte for the interrupt line, so irq 255 actually refers to 'not set'.
However, the 'irq' field for struct pci_dev is an integer, so the original meaning is lost, causing the system to assign an interrupt '255', which fails.

So we should _not_ assign an interrupt value here, and allow upper layers to fixup things.

This patch make PCI devices with MSI interrupts only (like the xhci device on certain HP laptops) work properly.

drivers/pci/probe.c +2 -1
index 6186f03..a2db887f 100644
@@ -923,7 +923,8 @@ static void pci_read_irq(struct pci_dev *dev)
 	dev->pin = irq;
 	if (irq)
 		pci_read_config_byte(dev, PCI_INTERRUPT_LINE, &irq);
-	dev->irq = irq;
+	if (irq < 255)
+		dev->irq = irq;
 }
 
 void set_pcie_port_type(struct pci_dev *pdev)
Tested-by on v1 1/1
#1 lore raw
Quoted 11 lines · Hannes Reinecke

The PCI config space reseves a byte for the interrupt line, so irq 255 actually refers to 'not set'.
However, the 'irq' field for struct pci_dev is an integer, so the original meaning is lost, causing the system to assign an interrupt '255', which fails.

So we should _not_ assign an interrupt value here, and allow upper layers to fixup things.

This patch make PCI devices with MSI interrupts only (like the xhci device on certain HP laptops) work properly.

Just tested and it works for me. Thank you.

Quoted 22 lines
drivers/pci/probe.c +2 -1
index 6186f03..a2db887f 100644
@@ -923,7 +923,8 @@ static void pci_read_irq(struct pci_dev *dev)
 	dev->pin = irq;
 	if (irq)
 		pci_read_config_byte(dev, PCI_INTERRUPT_LINE, &irq);
-	dev->irq = irq;
+	if (irq < 255)
+		dev->irq = irq;
 }
 
 void set_pcie_port_type(struct pci_dev *pdev)
Signature · 1 line
David Härdeman
#2 lore raw
In reply to Hannes Reinecke
Quoted 11 lines · Hannes Reinecke

The PCI config space reseves a byte for the interrupt line, so irq 255 actually refers to 'not set'.
However, the 'irq' field for struct pci_dev is an integer, so the original meaning is lost, causing the system to assign an interrupt '255', which fails.

So we should _not_ assign an interrupt value here, and allow upper layers to fixup things.

This patch make PCI devices with MSI interrupts only (like the xhci device on certain HP laptops) work properly.

looks like the bios does not provide _PRT for device in ACPI.

also according to PCI spec, BIOS *must* set interrupt line.

Quoted 27 lines
drivers/pci/probe.c +2 -2
index 6186f03..a2db887f 100644
@@ -923,7 +923,8 @@ static void pci_read_irq(struct pci_dev *dev)
        dev->pin = irq;
        if (irq)
                pci_read_config_byte(dev, PCI_INTERRUPT_LINE, &irq);
-       dev->irq = irq;
+       if (irq < 255)
+               dev->irq = irq;
 }
 
 void set_pcie_port_type(struct pci_dev *pdev)
--

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

#3 lore raw
Quoted 7 lines · Yinghai Lu
Quoted 6 lines · Hannes Reinecke

The PCI config space reseves a byte for the interrupt line, so irq 255 actually refers to 'not set'.
However, the 'irq' field for struct pci_dev is an integer, so the original meaning is lost, causing the system to assign an interrupt '255', which fails.

Quoted 8 lines
Quoted 5 lines

So we should _not_ assign an interrupt value here, and allow upper layers to fixup things.

This patch make PCI devices with MSI interrupts only (like the xhci device on certain HP laptops) work properly.

looks like the bios does not provide _PRT for device in ACPI.

Correct.

Quoted 2 lines

also according to PCI spec, BIOS *must* set interrupt line.

Apparently this device is meant to use MSI _only_ so the BIOS developer didn't feel the need to assign an INTx here.

According to PCI-3.0, section 6.8 (Message Signalled Interrupts):

Quoted 8 lines

It is recommended that devices implement interrupt pins to
provide compatibility in systems that do not support MSI
(devices default to interrupt pins). However, it is expected
that the need for interrupt pins will diminish over time.
Devices that do not support interrupt pins due to pin
constraints (rely on polling for device service) may implement
messages to increase performance without adding additional pins. > Therefore, system configuration software must not assume that a message capable device has an interrupt pin.

Which sounds to me as if the implementation is valid... And in either case, I've added the relevant details plus patch to bnc#52591.
Including ACPI dump, so you can check for yourself.

And correct me if I'm wrong, of course :-)

Cheers,

Hannes

Signature · 4 lines
Dr. Hannes Reinecke		      zSeries & Storage
[email protected]			      +49 911 74053 688
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: J. Hawn, J. Guild, F. Imendörffer, HRB 16746 (AG Nürnberg)
#4 lore raw
Quoted 15 lines · Hannes Reinecke
Quoted 1 line

Apparently this device is meant to use MSI _only_ so the BIOS developer didn't feel the need to assign an INTx here.

According to PCI-3.0, section 6.8 (Message Signalled Interrupts):

Quoted 9 lines

It is recommended that devices implement interrupt pins to provide compatibility in systems that do not support MSI (devices default to interrupt pins). However, it is expected that the need for interrupt pins will diminish over time. Devices that do not support interrupt pins due to pin constraints (rely on polling for device service) may implement messages to increase performance without adding additional pins. > Therefore, system configuration software must not assume that a message capable device has an interrupt pin.

Quoted 1 line

Which sounds to me as if the implementation is valid...

it seems you mess pin with interrupt line.

current code:

        unsigned char irq;

        pci_read_config_byte(dev, PCI_INTERRUPT_PIN, &irq);
        dev->pin = irq;
        if (irq)
                pci_read_config_byte(dev, PCI_INTERRUPT_LINE, &irq);
        dev->irq = irq;

so if the device does not have interrupt pin implemented, pin should be zero. and pin and irq in dev should
be all 0.

Thanks

Yinghai

#5 lore raw
Quoted 29 lines · Yinghai Lu
Quoted 15 lines · Hannes Reinecke
Quoted 1 line

Apparently this device is meant to use MSI _only_ so the BIOS developer didn't feel the need to assign an INTx here.

According to PCI-3.0, section 6.8 (Message Signalled Interrupts):

Quoted 9 lines

It is recommended that devices implement interrupt pins to provide compatibility in systems that do not support MSI (devices default to interrupt pins). However, it is expected that the need for interrupt pins will diminish over time. Devices that do not support interrupt pins due to pin constraints (rely on polling for device service) may implement messages to increase performance without adding additional pins. > Therefore, system configuration software must not assume that a message capable device has an interrupt pin.

Quoted 1 line

Which sounds to me as if the implementation is valid...

it seems you mess pin with interrupt line.

current code:

         unsigned char irq;

         pci_read_config_byte(dev, PCI_INTERRUPT_PIN, &irq);
         dev->pin = irq;
         if (irq)
                 pci_read_config_byte(dev, PCI_INTERRUPT_LINE, &irq);
         dev->irq = irq;
Quoted 4 lines

so if the device does not have interrupt pin implemented, pin should be zero. and pin and irq in dev should
be all 0.

But the device _has_ an interrupt pin implemented. The whole point here is that the interrupt line is _NOT_ zero.

00:14.0 USB controller [0c03]: Intel Corporation 7 Series/C210 Series Chipset Family USB xHCI Host Controller [8086:1e31] (rev 04) (prog-if 30 [XHCI])
Subsystem: Hewlett-Packard Company Device [103c:179b]
Control: I/O- Mem+ BusMaster- SpecCycle- MemWINV- VGASnoop- ParErr-
Stepping- SERR- FastB2B- DisINTx-
Status: Cap+ 66MHz- UDF- FastB2B+ ParErr- DEVSEL=medium >TAbort-
<TAbort- <MAbort- >SERR- <PERR- INTx-
Interrupt: pin A routed to IRQ 255
Region 0: Memory at d4720000 (64-bit, non-prefetchable) [size=64K]
Capabilities: [70] Power Management version 2
Flags: PMEClk- DSI- D1- D2- AuxCurrent=375mA
PME(D0-,D1-,D2-,D3hot+,D3cold+)

		Status: D0 NoSoftRst+ PME-Enable- DSel=0 DScale=0 PME-
	Capabilities: [80] MSI: Enable- Count=1/8 Maskable- 64bit+
		Address: 0000000000000000  Data: 0000

So at one point we have to decide that ->irq is not valid, despite it being not set to zero. An alternative fix would be this:

drivers/acpi/pci_irq.c +1 -0
index 68a921d..4a480cb 100644
@@ -469,6 +469,7 @@ int acpi_pci_irq_enable(struct pci_dev *dev)
                } else {
                        dev_warn(&dev->dev, "PCI INT %c: no GSI\n",
                                 pin_name(pin));
+                       dev->irq = 0;
                }
                return 0;
        }

Which probably is a better solution, as here ->irq is _definitely_ not valid, so we should reset it to '0' to avoid confusion on upper layers.

Cheers,

Hannes

Signature · 4 lines
Dr. Hannes Reinecke		      zSeries & Storage
[email protected]			      +49 911 74053 688
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: J. Hawn, J. Guild, F. Imendörffer, HRB 16746 (AG Nürnberg)
#6 lore raw
Quoted 19 lines · Hannes Reinecke
Quoted 11 lines · Yinghai Lu

it seems you mess pin with interrupt line.

current code:

        unsigned char irq;

        pci_read_config_byte(dev, PCI_INTERRUPT_PIN, &irq);
        dev->pin = irq;
        if (irq)
                pci_read_config_byte(dev, PCI_INTERRUPT_LINE, &irq);
        dev->irq = irq;
Quoted 4 lines

so if the device does not have interrupt pin implemented, pin should be zero. and pin and irq in dev should
be all 0.

But the device _has_ an interrupt pin implemented. The whole point here is that the interrupt line is _NOT_ zero.

...

Quoted 18 lines

So at one point we have to decide that ->irq is not valid, despite it being not set to zero.
An alternative fix would be this:

drivers/acpi/pci_irq.c +1 -0
index 68a921d..4a480cb 100644
@@ -469,6 +469,7 @@ int acpi_pci_irq_enable(struct pci_dev *dev)
               } else {
                       dev_warn(&dev->dev, "PCI INT %c: no GSI\n",
                                pin_name(pin));
+                       dev->irq = 0;
               }
               return 0;
       }
Quoted 4 lines

Which probably is a better solution, as here ->irq is _definitely_ not valid, so we should reset it to '0' to avoid confusion on upper layers.

Is there any agreement on how to proceed?

Signature · 1 line
David Härdeman
#7 lore raw
Quoted 39 lines · David Härdeman
Quoted 19 lines · Hannes Reinecke
Quoted 11 lines · Yinghai Lu

it seems you mess pin with interrupt line.

current code:

         unsigned char irq;

         pci_read_config_byte(dev, PCI_INTERRUPT_PIN, &irq);
         dev->pin = irq;
         if (irq)
                 pci_read_config_byte(dev, PCI_INTERRUPT_LINE, &irq);
         dev->irq = irq;
Quoted 4 lines

so if the device does not have interrupt pin implemented, pin should be zero. and pin and irq in dev should
be all 0.

But the device _has_ an interrupt pin implemented. The whole point here is that the interrupt line is _NOT_ zero.

...

Quoted 18 lines

So at one point we have to decide that ->irq is not valid, despite it being not set to zero.
An alternative fix would be this:

drivers/acpi/pci_irq.c +1 -0
index 68a921d..4a480cb 100644
@@ -469,6 +469,7 @@ int acpi_pci_irq_enable(struct pci_dev *dev)
                } else {
                        dev_warn(&dev->dev, "PCI INT %c: no GSI\n",
                                 pin_name(pin));
+                       dev->irq = 0;
                }
                return 0;
        }
Quoted 7 lines
Quoted 4 lines

Which probably is a better solution, as here ->irq is _definitely_ not valid, so we should reset it to '0' to avoid confusion on upper layers.

Is there any agreement on how to proceed?

I would actually prefer the second solution, as the ACPI code gives
some better guarantees here. With the original solution it _might_ be that on non-ACPI systems an interrupt 255 is valid, so it might incur unwanted regressions.

However, for an ACPI system we only have the two choices, assigning an interrupt via ACPI tables or use a default GSI value. If both failed the interrupt definitely is not valid and can safely be reset to 0.

But this would need a formal ACK from the ACPI gods ... Len? Rafael?

Cheers,

Hannes

Signature · 4 lines
Dr. Hannes Reinecke		      zSeries & Storage
[email protected]			      +49 911 74053 688
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: J. Hawn, J. Guild, F. Imendörffer, HRB 16746 (AG Nürnberg)
maintainer
#8 lore raw
In reply to Hannes Reinecke

[+cc Andy]

Quoted 77 lines · Hannes Reinecke
Quoted 34 lines · Yinghai Lu
Quoted 18 lines · Hannes Reinecke
Quoted 2 lines

Apparently this device is meant to use MSI _only_ so the BIOS developer didn't feel the need to assign an INTx here.

According to PCI-3.0, section 6.8 (Message Signalled Interrupts):

Quoted 10 lines

It is recommended that devices implement interrupt pins to provide compatibility in systems that do not support MSI (devices default to interrupt pins). However, it is expected that the need for interrupt pins will diminish over time. Devices that do not support interrupt pins due to pin constraints (rely on polling for device service) may implement messages to increase performance without adding additional pins. > Therefore, system configuration software must not assume that a message capable device has an interrupt pin.

Quoted 1 line

Which sounds to me as if the implementation is valid...

it seems you mess pin with interrupt line.

current code:

         unsigned char irq;

         pci_read_config_byte(dev, PCI_INTERRUPT_PIN, &irq);
         dev->pin = irq;
         if (irq)
                 pci_read_config_byte(dev, PCI_INTERRUPT_LINE, &irq);
         dev->irq = irq;
Quoted 5 lines

so if the device does not have interrupt pin implemented, pin should be zero.
and pin and irq in dev should
be all 0.

But the device _has_ an interrupt pin implemented. The whole point here is that the interrupt line is _NOT_ zero.

00:14.0 USB controller [0c03]: Intel Corporation 7 Series/C210 Series Chipset Family USB xHCI Host Controller [8086:1e31] (rev 04) (prog-if 30 [XHCI])
Subsystem: Hewlett-Packard Company Device [103c:179b]
Control: I/O- Mem+ BusMaster- SpecCycle- MemWINV- VGASnoop- ParErr-
Stepping- SERR- FastB2B- DisINTx-
Status: Cap+ 66MHz- UDF- FastB2B+ ParErr- DEVSEL=medium >TAbort-
<TAbort- <MAbort- >SERR- <PERR- INTx-
Interrupt: pin A routed to IRQ 255
Region 0: Memory at d4720000 (64-bit, non-prefetchable) [size=64K]
Capabilities: [70] Power Management version 2
Flags: PMEClk- DSI- D1- D2- AuxCurrent=375mA
PME(D0-,D1-,D2-,D3hot+,D3cold+)

                Status: D0 NoSoftRst+ PME-Enable- DSel=0 DScale=0 PME-
        Capabilities: [80] MSI: Enable- Count=1/8 Maskable- 64bit+
                Address: 0000000000000000  Data: 0000

So at one point we have to decide that ->irq is not valid, despite it being not set to zero.
An alternative fix would be this:

drivers/acpi/pci_irq.c +1 -0
index 68a921d..4a480cb 100644
@@ -469,6 +469,7 @@ int acpi_pci_irq_enable(struct pci_dev *dev)
                } else {
                        dev_warn(&dev->dev, "PCI INT %c: no GSI\n",
                                 pin_name(pin));
+                       dev->irq = 0;
                }
                return 0;
        }
Quoted 3 lines

Which probably is a better solution, as here ->irq is _definitely_ not valid, so we should reset it to '0' to avoid confusion on upper layers.

I didn't like the pci_read_irq() change because the PCI spec doesn't say anything about any PCI_INTERRUPT_LINE values being invalid.

I like this solution better, but I still don't quite understand it. We have the following code in acpi_pci_irq_enable(). We have previously tried to look up "gsi," but the _PRT doesn't mention this device, so we have "gsi == -1" at this point:

        /*
         * No IRQ known to the ACPI subsystem - maybe the BIOS /
         * driver reported one, then use it. Exit in any case.
         */
        if (gsi < 0) {
                u32 dev_gsi;
                /* Interrupt Line values above 0xF are forbidden */
                if (dev->irq > 0 && (dev->irq <= 0xF) &&
                    (acpi_isa_irq_to_gsi(dev->irq, &dev_gsi) == 0)) {
                        dev_warn(&dev->dev, "PCI INT %c: no GSI -
using ISA IRQ %d\n",
                                 pin_name(pin), dev->irq);
                        acpi_register_gsi(&dev->dev, dev_gsi,
                                          ACPI_LEVEL_SENSITIVE,
                                          ACPI_ACTIVE_LOW);
                } else {
                        dev_warn(&dev->dev, "PCI INT %c: no GSI\n",
                                 pin_name(pin));
                }

                return 0;
        }

1) I don't know where the restriction of 0x1-0xF came from. Presumably this value of dev->irq came from PCI_INTERRUPT_LINE, and I don't know what forbids values > 0xF. The test was added by Andy Grover in the attached commit. This is ancient history; probably Andy doesn't remember either :)

2) The proposed change (setting "dev->irq = 0" when we didn't find anything in the _PRT and dev->irq > 0xF) throws away some information, and I fear it could break systems. For example, what would happen if a system put an IOAPIC pin number in a device's PCI_INTERRUPT_LINE and omitted the device from _PRT? Is it possible the device would still work as-is (with acpi_pci_irq_enable() doing nothing), but would break if we set dev->irq = 0?

3) I don't understand why the xhci init fails in the first place. It looks like the "request interrupt 255 failed" message is from xhci_try_enable_msi(), but that function tries to enable MSI-X, then MSI, then falls back to legacy interrupts, where we get the error. But the device supports MSI, so I don't know why we even fall back to trying legacy interrupts. Hannes, do you have any insight into that? Obviously I'm missing something here.

Bjorn

#9 lore raw
Quoted 129 lines · Bjorn Helgaas

[+cc Andy]

Quoted 77 lines · Hannes Reinecke
Quoted 34 lines · Yinghai Lu
Quoted 18 lines · Hannes Reinecke
Quoted 2 lines

Apparently this device is meant to use MSI _only_ so the BIOS developer didn't feel the need to assign an INTx here.

According to PCI-3.0, section 6.8 (Message Signalled Interrupts):

Quoted 10 lines

It is recommended that devices implement interrupt pins to provide compatibility in systems that do not support MSI (devices default to interrupt pins). However, it is expected that the need for interrupt pins will diminish over time. Devices that do not support interrupt pins due to pin constraints (rely on polling for device service) may implement messages to increase performance without adding additional pins. > Therefore, system configuration software must not assume that a message capable device has an interrupt pin.

Quoted 1 line

Which sounds to me as if the implementation is valid...

it seems you mess pin with interrupt line.

current code:

          unsigned char irq;

          pci_read_config_byte(dev, PCI_INTERRUPT_PIN, &irq);
          dev->pin = irq;
          if (irq)
                  pci_read_config_byte(dev, PCI_INTERRUPT_LINE, &irq);
          dev->irq = irq;
Quoted 5 lines

so if the device does not have interrupt pin implemented, pin should be zero.
and pin and irq in dev should
be all 0.

But the device _has_ an interrupt pin implemented. The whole point here is that the interrupt line is _NOT_ zero.

00:14.0 USB controller [0c03]: Intel Corporation 7 Series/C210 Series Chipset Family USB xHCI Host Controller [8086:1e31] (rev 04) (prog-if 30 [XHCI])
Subsystem: Hewlett-Packard Company Device [103c:179b]
Control: I/O- Mem+ BusMaster- SpecCycle- MemWINV- VGASnoop- ParErr-
Stepping- SERR- FastB2B- DisINTx-
Status: Cap+ 66MHz- UDF- FastB2B+ ParErr- DEVSEL=medium >TAbort-
<TAbort- <MAbort- >SERR- <PERR- INTx-
Interrupt: pin A routed to IRQ 255
Region 0: Memory at d4720000 (64-bit, non-prefetchable) [size=64K]
Capabilities: [70] Power Management version 2
Flags: PMEClk- DSI- D1- D2- AuxCurrent=375mA
PME(D0-,D1-,D2-,D3hot+,D3cold+)

                 Status: D0 NoSoftRst+ PME-Enable- DSel=0 DScale=0 PME-
         Capabilities: [80] MSI: Enable- Count=1/8 Maskable- 64bit+
                 Address: 0000000000000000  Data: 0000

So at one point we have to decide that ->irq is not valid, despite it being not set to zero.
An alternative fix would be this:

drivers/acpi/pci_irq.c +1 -0
index 68a921d..4a480cb 100644
@@ -469,6 +469,7 @@ int acpi_pci_irq_enable(struct pci_dev *dev)
                 } else {
                         dev_warn(&dev->dev, "PCI INT %c: no GSI\n",
                                  pin_name(pin));
+                       dev->irq = 0;
                 }
                 return 0;
         }
Quoted 3 lines

Which probably is a better solution, as here ->irq is _definitely_ not valid, so we should reset it to '0' to avoid confusion on upper layers.

I didn't like the pci_read_irq() change because the PCI spec doesn't say anything about any PCI_INTERRUPT_LINE values being invalid.

I like this solution better, but I still don't quite understand it. We have the following code in acpi_pci_irq_enable(). We have previously tried to look up "gsi," but the _PRT doesn't mention this device, so we have "gsi == -1" at this point:

         /*
          * No IRQ known to the ACPI subsystem - maybe the BIOS /
          * driver reported one, then use it. Exit in any case.
          */
         if (gsi < 0) {
                 u32 dev_gsi;
                 /* Interrupt Line values above 0xF are forbidden */
                 if (dev->irq > 0 && (dev->irq <= 0xF) &&
                     (acpi_isa_irq_to_gsi(dev->irq, &dev_gsi) == 0)) {
                         dev_warn(&dev->dev, "PCI INT %c: no GSI -
using ISA IRQ %d\n",
                                  pin_name(pin), dev->irq);
                         acpi_register_gsi(&dev->dev, dev_gsi,
                                           ACPI_LEVEL_SENSITIVE,
                                           ACPI_ACTIVE_LOW);
                 } else {
                         dev_warn(&dev->dev, "PCI INT %c: no GSI\n",
                                  pin_name(pin));
                 }

                 return 0;
         }

1) I don't know where the restriction of 0x1-0xF came from. Presumably this value of dev->irq came from PCI_INTERRUPT_LINE, and I don't know what forbids values > 0xF. The test was added by Andy Grover in the attached commit. This is ancient history; probably Andy doesn't remember either :)

2) The proposed change (setting "dev->irq = 0" when we didn't find anything in the _PRT and dev->irq > 0xF) throws away some information, and I fear it could break systems. For example, what would happen if a system put an IOAPIC pin number in a device's PCI_INTERRUPT_LINE and omitted the device from _PRT? Is it possible the device would still work as-is (with acpi_pci_irq_enable() doing nothing), but would break if we set dev->irq = 0?

Quoted 8 lines

3) I don't understand why the xhci init fails in the first place. It looks like the "request interrupt 255 failed" message is from xhci_try_enable_msi(), but that function tries to enable MSI-X, then MSI, then falls back to legacy interrupts, where we get the error. But the device supports MSI, so I don't know why we even fall back to trying legacy interrupts. Hannes, do you have any insight into that? Obviously I'm missing something here.

Right.

It's infact a quirk with the USB HCD setup.

drivers/usb/core/hcd.c calls request_irq for the legacy interrupt line, before drivers/usb/host/xhci.c enables MSI-X.

So the correct thing would be to skip the request_irq() function in drivers/usb/core/hcd.c.

I'll be drafting up a patch.

Cheers,

Hannes

Signature · 4 lines
Dr. Hannes Reinecke		      zSeries & Storage
[email protected]			      +49 911 74053 688
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: J. Hawn, J. Guild, F. Imendörffer, HRB 16746 (AG Nürnberg)
#10 lore raw
In reply to Bjorn Helgaas
Quoted 115 lines · Bjorn Helgaas

[+cc Andy]

Quoted 77 lines · Hannes Reinecke
Quoted 34 lines · Yinghai Lu
Quoted 18 lines · Hannes Reinecke
Quoted 2 lines

Apparently this device is meant to use MSI _only_ so the BIOS developer didn't feel the need to assign an INTx here.

According to PCI-3.0, section 6.8 (Message Signalled Interrupts):

Quoted 10 lines

It is recommended that devices implement interrupt pins to provide compatibility in systems that do not support MSI (devices default to interrupt pins). However, it is expected that the need for interrupt pins will diminish over time. Devices that do not support interrupt pins due to pin constraints (rely on polling for device service) may implement messages to increase performance without adding additional pins. > Therefore, system configuration software must not assume that a message capable device has an interrupt pin.

Quoted 1 line

Which sounds to me as if the implementation is valid...

it seems you mess pin with interrupt line.

current code:

          unsigned char irq;

          pci_read_config_byte(dev, PCI_INTERRUPT_PIN, &irq);
          dev->pin = irq;
          if (irq)
                  pci_read_config_byte(dev, PCI_INTERRUPT_LINE, &irq);
          dev->irq = irq;
Quoted 5 lines

so if the device does not have interrupt pin implemented, pin should be zero.
and pin and irq in dev should
be all 0.

But the device _has_ an interrupt pin implemented. The whole point here is that the interrupt line is _NOT_ zero.

00:14.0 USB controller [0c03]: Intel Corporation 7 Series/C210 Series Chipset Family USB xHCI Host Controller [8086:1e31] (rev 04) (prog-if 30 [XHCI])
Subsystem: Hewlett-Packard Company Device [103c:179b]
Control: I/O- Mem+ BusMaster- SpecCycle- MemWINV- VGASnoop- ParErr-
Stepping- SERR- FastB2B- DisINTx-
Status: Cap+ 66MHz- UDF- FastB2B+ ParErr- DEVSEL=medium >TAbort-
<TAbort- <MAbort- >SERR- <PERR- INTx-
Interrupt: pin A routed to IRQ 255
Region 0: Memory at d4720000 (64-bit, non-prefetchable) [size=64K]
Capabilities: [70] Power Management version 2
Flags: PMEClk- DSI- D1- D2- AuxCurrent=375mA
PME(D0-,D1-,D2-,D3hot+,D3cold+)

                 Status: D0 NoSoftRst+ PME-Enable- DSel=0 DScale=0 PME-
         Capabilities: [80] MSI: Enable- Count=1/8 Maskable- 64bit+
                 Address: 0000000000000000  Data: 0000

So at one point we have to decide that ->irq is not valid, despite it being not set to zero.
An alternative fix would be this:

drivers/acpi/pci_irq.c +1 -0
index 68a921d..4a480cb 100644
@@ -469,6 +469,7 @@ int acpi_pci_irq_enable(struct pci_dev *dev)
                 } else {
                         dev_warn(&dev->dev, "PCI INT %c: no GSI\n",
                                  pin_name(pin));
+                       dev->irq = 0;
                 }
                 return 0;
         }
Quoted 3 lines

Which probably is a better solution, as here ->irq is _definitely_ not valid, so we should reset it to '0' to avoid confusion on upper layers.

I didn't like the pci_read_irq() change because the PCI spec doesn't say anything about any PCI_INTERRUPT_LINE values being invalid.

I like this solution better, but I still don't quite understand it. We have the following code in acpi_pci_irq_enable(). We have previously tried to look up "gsi," but the _PRT doesn't mention this device, so we have "gsi == -1" at this point:

         /*
          * No IRQ known to the ACPI subsystem - maybe the BIOS /
          * driver reported one, then use it. Exit in any case.
          */
         if (gsi < 0) {
                 u32 dev_gsi;
                 /* Interrupt Line values above 0xF are forbidden */
                 if (dev->irq > 0 && (dev->irq <= 0xF) &&
                     (acpi_isa_irq_to_gsi(dev->irq, &dev_gsi) == 0)) {
                         dev_warn(&dev->dev, "PCI INT %c: no GSI -
using ISA IRQ %d\n",
                                  pin_name(pin), dev->irq);
                         acpi_register_gsi(&dev->dev, dev_gsi,
                                           ACPI_LEVEL_SENSITIVE,
                                           ACPI_ACTIVE_LOW);
                 } else {
                         dev_warn(&dev->dev, "PCI INT %c: no GSI\n",
                                  pin_name(pin));
                 }

                 return 0;
         }
Quoted 6 lines

1) I don't know where the restriction of 0x1-0xF came from. Presumably this value of dev->irq came from PCI_INTERRUPT_LINE, and I don't know what forbids values > 0xF. The test was added by Andy Grover in the attached commit. This is ancient history; probably Andy doesn't remember either :)

This is most likely due to ISA compability. Cf ACPI 4.0, section 5.2.12.4 Platforms with APIC and Dual 8259 Support:

Quoted 20 lines

Systems that support both APIC and dual 8259 interrupt models must map global system interrupts 0-15 to the 8259 IRQs 0-15, except where Interrupt Source Overrides are provided (see section 5.2.10.8, “Interrupt Source Overrides”). This means that I/O APIC interrupt inputs 0-15 must be mapped to global system interrupts 0-15 and have identical sources as the 8259 IRQs 0-15 unless overrides are used. This allows a platform to support OSPM implementations that use the APIC model as well as OSPM implementations that use the 8259 model (OSPM will only use one model; it will not mix models).
When OSPM supports the 8259 model, it will assume that all interrupt descriptors reporting global system interrupts 0-15 correspond to 8259 IRQs. In the 8259 model all global system interrupts greater than 15 are ignored. If OSPM implements APIC support, it will enable the APIC as described by the APIC specification and will use all reported global system interrupts that fall within the limits of the interrupt inputs defined by the I/O APIC structures. For more information on hardware resource configuration see section 6, “Configuration.”

Quoted 8 lines

2) The proposed change (setting "dev->irq = 0" when we didn't find anything in the _PRT and dev->irq > 0xF) throws away some information, and I fear it could break systems. For example, what would happen if a system put an IOAPIC pin number in a device's PCI_INTERRUPT_LINE and omitted the device from _PRT? Is it possible the device would still work as-is (with acpi_pci_irq_enable() doing nothing), but would break if we set dev->irq = 0?

How so? It would still be required to get an interrupt from somewhere, and if ACPI doesn't know about it where should the information come from?

Quoted 8 lines

3) I don't understand why the xhci init fails in the first place. It looks like the "request interrupt 255 failed" message is from xhci_try_enable_msi(), but that function tries to enable MSI-X, then MSI, then falls back to legacy interrupts, where we get the error. But the device supports MSI, so I don't know why we even fall back to trying legacy interrupts. Hannes, do you have any insight into that? Obviously I'm missing something here.

Hehe. Due to overly clever design.
xhci actually sets up interrupts _twice_, once per request_irq() in the generic code and a second time during xhci_run. But as the first call fails it'll never ever run the second part.

I'll be sending a patch.

Cheers,

Hannes

Signature · 4 lines
Dr. Hannes Reinecke		      zSeries & Storage
[email protected]			      +49 911 74053 688
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: J. Hawn, J. Guild, F. Imendörffer, HRB 16746 (AG Nürnberg)
#11 lore raw
Quoted 11 lines · Hannes Reinecke
Quoted 10 lines · Bjorn Helgaas

[+cc Andy]

3) I don't understand why the xhci init fails in the first place. It looks like the "request interrupt 255 failed" message is from xhci_try_enable_msi(), but that function tries to enable MSI-X, then MSI, then falls back to legacy interrupts, where we get the error. But the device supports MSI, so I don't know why we even fall back to trying legacy interrupts. Hannes, do you have any insight into that? Obviously I'm missing something here.

Quoted 6 lines

Hehe. Due to overly clever design.
xhci actually sets up interrupts _twice_, once per request_irq() in the generic code and a second time during xhci_run.
But as the first call fails it'll never ever run the second part.

I'll be sending a patch.

Something like this?

http://marc.info/?l=linux-usb&m=132972894117916&w=2

(Apparently we had this issue around the same time last year, but we thought the BIOS bug had been resolved.)

Sarah Sharp

maintainer
#12 lore raw
Quoted 19 lines · Sarah Sharp
Quoted 11 lines · Hannes Reinecke
Quoted 10 lines · Bjorn Helgaas

[+cc Andy]

3) I don't understand why the xhci init fails in the first place. It looks like the "request interrupt 255 failed" message is from xhci_try_enable_msi(), but that function tries to enable MSI-X, then MSI, then falls back to legacy interrupts, where we get the error. But the device supports MSI, so I don't know why we even fall back to trying legacy interrupts. Hannes, do you have any insight into that? Obviously I'm missing something here.

Quoted 6 lines

Hehe. Due to overly clever design.
xhci actually sets up interrupts _twice_, once per request_irq() in the generic code and a second time during xhci_run.
But as the first call fails it'll never ever run the second part.

I'll be sending a patch.

Quoted 8 lines

Something like this?

http://marc.info/?l=linux-usb&m=132972894117916&w=2

(Apparently we had this issue around the same time last year, but we thought the BIOS bug had been resolved.)

Sarah Sharp

Where are we at with this? I don't see Sarah's patch in the tree, and I haven't applied any changes, so my guess is this is still broken.

Bjorn

maintainer
#14 lore raw
Quoted 8 lines · Yinghai Lu
Quoted 2 lines · Bjorn Helgaas

Where are we at with this? I don't see Sarah's patch in the tree, and I haven't applied any changes, so my guess is this is still broken.

Yes, the fix is in the linus tree.

https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=00eed9c814cb8f281be6f0f5d8f45025dc0a97eb

Thanks, Yinghai.

Frederik, can you confirm that
https://bugzilla.kernel.org/show_bug.cgi?id=52591 is fixed by the change above?

Bjorn

maintainer
#15 lore raw
In reply to Hannes Reinecke
Quoted 24 lines · Hannes Reinecke

The PCI config space reseves a byte for the interrupt line, so irq 255 actually refers to 'not set'.
However, the 'irq' field for struct pci_dev is an integer, so the original meaning is lost, causing the system to assign an interrupt '255', which fails.

So we should _not_ assign an interrupt value here, and allow upper layers to fixup things.

This patch make PCI devices with MSI interrupts only (like the xhci device on certain HP laptops) work properly.

drivers/pci/probe.c +0 -0
index 6186f03..a2db887f 100644

drivers/pci/probe.c static void pci_read_irq(struct pci_dev *dev) · line 930

Quoted 9 lines
drivers/pci/probe.c +2 -1
@@ -923,7 +923,8 @@ static void pci_read_irq(struct pci_dev *dev)
        dev->pin = irq;
        if (irq)
                pci_read_config_byte(dev, PCI_INTERRUPT_LINE, &irq);
-       dev->irq = irq;
+       if (irq < 255)
+               dev->irq = irq;
 }
 
 void set_pcie_port_type(struct pci_dev *pdev)

Is there a bugzilla or other URL with more information (problem description, hardware involved, dmesg logs, acpidump, etc)?

Bjorn

#16 lore raw
Quoted 33 lines · Bjorn Helgaas
Quoted 32 lines · Hannes Reinecke

The PCI config space reseves a byte for the interrupt line, so irq 255 actually refers to 'not set'.
However, the 'irq' field for struct pci_dev is an integer, so the original meaning is lost, causing the system to assign an interrupt '255', which fails.

So we should _not_ assign an interrupt value here, and allow upper layers to fixup things.

This patch make PCI devices with MSI interrupts only (like the xhci device on certain HP laptops) work properly.

drivers/pci/probe.c +2 -1
index 6186f03..a2db887f 100644
@@ -923,7 +923,8 @@ static void pci_read_irq(struct pci_dev *dev)
        dev->pin = irq;
        if (irq)
                pci_read_config_byte(dev, PCI_INTERRUPT_LINE, &irq);
-       dev->irq = irq;
+       if (irq < 255)
+               dev->irq = irq;
 }
 
Quoted 4 lines

drivers/pci/probe.c static void pci_read_irq(struct pci_dev *dev) · line 930

Quoted 1 line
drivers/pci/probe.c +0 -0
 void set_pcie_port_type(struct pci_dev *pdev)

Is there a bugzilla or other URL with more information (problem description, hardware involved, dmesg logs, acpidump, etc)?

https://bugzilla.kernel.org/show_bug.cgi?id=52591
https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1072918

Basically it looks like most HP Probook/Elitebook of the Ivy Bridge generation are affected.

Signature · 1 line
Frederik Himpe