Skip to contentSkip to conversation

PCI/MSI: Clarify the IRQ sysfs ABI for PCI devices

PATCH Superseded Barry Song 3 participants 19 messages Aug 20, 2021 – Aug 25, 2021 PCI

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Barry Song

author
#0 lore raw

v2 cover PCI/MSI: Clarify the IRQ sysfs ABI for PCI devices

From: Barry Song <[email protected]>

/sys/bus/pci/devices/.../irq has been there for many years but it has never been documented. This patch is trying to document it. Plus, irq ABI is very confusing at this moment especially for MSI and MSI-x cases. MSI sets irq to the first number in the vector, but MSI-X does nothing for this though it saves default_irq in msix_setup_entries(). Weird the saved default_irq for MSI-X is never used in pci_msix_shutdown(), which is quite different with pci_msi_shutdown(). Thus, this patch also moves to show the first IRQ number which is from the first msi_entry for MSI-X. Hopefully, this can make irq ABI more clear and more consistent.

-v2:
- split into two patches according to Bjorn's comments;
- Add Greg's Acked-by, thanks for reviewing!
-v1:
https://lore.kernel.org/lkml/[email protected]/#t

Barry Song (2):
PCI/MSI: Fix the confusing IRQ sysfs ABI for MSI-X
Documentation: ABI: sysfs-bus-pci: Add description for IRQ entry

 Documentation/ABI/testing/sysfs-bus-pci | 8 ++++++++
 drivers/pci/msi.c                       | 6 ++++++
 2 files changed, 14 insertions(+)
Signature · 1 line
1.8.3.1
author
#1 lore raw

PATCH v2 1/2 PCI/MSI: Fix the confusing IRQ sysfs ABI for MSI-X

From: Barry Song <[email protected]>

/sys/bus/pci/devices/.../irq sysfs ABI is very confusing at this moment especially for MSI-X cases. While MSI sets IRQ to the first number in the vector, MSI-X does nothing for this though it saves default_irq in msix_setup_entries(). Weird the saved default_irq for MSI-X is never used in pci_msix_shutdown(), which is quite different with pci_msi_shutdown(). Thus, this patch moves to show the first IRQ number which is from the first msi_entry for MSI-X. Hopefully, this can make IRQ ABI more clear and more consistent.

---
 drivers/pci/msi.c | 6 ++++++
 1 file changed, 6 insertions(+)
drivers/pci/msi.c +6 -0
index 9232255..6bbf81b 100644
@@ -771,6 +771,7 @@ static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries,
 	int ret;
 	u16 control;
 	void __iomem *base;
+	struct msi_desc *desc;
 
 	/* Ensure MSI-X is disabled while it is set up */
 	pci_msix_clear_and_set_ctrl(dev, PCI_MSIX_FLAGS_ENABLE, 0);
@@ -814,6 +815,10 @@ static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries,
 	pci_msix_clear_and_set_ctrl(dev, PCI_MSIX_FLAGS_MASKALL, 0);
 
 	pcibios_free_irq(dev);
+
+	desc = first_pci_msi_entry(dev);
+	dev->irq = desc->irq;
+
 	return 0;
 
 out_avail:
@@ -1024,6 +1029,7 @@ static void pci_msix_shutdown(struct pci_dev *dev)
 	pci_msix_clear_and_set_ctrl(dev, PCI_MSIX_FLAGS_ENABLE, 0);
 	pci_intx_for_msi(dev, 1);
 	dev->msix_enabled = 0;
+	dev->irq = entry->msi_attrib.default_irq;
 	pcibios_alloc_irq(dev);
 }
 
Signature · 1 line
1.8.3.1
#2 lore raw

Re: [PATCH v2 1/2] PCI/MSI: Fix the confusing IRQ sysfs ABI for MSI-X

[+cc Thomas, Marc]

Quoted 4 lines · Barry Song

From: Barry Song <[email protected]>

/sys/bus/pci/devices/.../irq sysfs ABI is very confusing at this moment especially for MSI-X cases.

AFAICT this patch *only* affects MSI-X. So are you saying the sysfs ABI is fine for MSI but confusing for MSI-X?

Quoted 27 lines

While MSI sets IRQ to the first
number in the vector, MSI-X does nothing for this though it saves default_irq in msix_setup_entries(). Weird the saved default_irq for MSI-X is never used in pci_msix_shutdown(), which is quite different with pci_msi_shutdown(). Thus, this patch moves to show the first IRQ number which is from the first msi_entry for MSI-X. Hopefully, this can make IRQ ABI more clear and more consistent.

---
 drivers/pci/msi.c | 6 ++++++
 1 file changed, 6 insertions(+)
drivers/pci/msi.c +1 -0
index 9232255..6bbf81b 100644
@@ -771,6 +771,7 @@ static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries,
 	int ret;
 	u16 control;
 	void __iomem *base;
+	struct msi_desc *desc;
 
 	/* Ensure MSI-X is disabled while it is set up */
 	pci_msix_clear_and_set_ctrl(dev, PCI_MSIX_FLAGS_ENABLE, 0);

drivers/pci/msi.c static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries, · line 820

Quoted 6 lines
drivers/pci/msi.c +3 -0
@@ -814,6 +815,10 @@ static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries,
 	pci_msix_clear_and_set_ctrl(dev, PCI_MSIX_FLAGS_MASKALL, 0);
 
 	pcibios_free_irq(dev);
+
+	desc = first_pci_msi_entry(dev);
+	dev->irq = desc->irq;

This change is not primarily about sysfs. This is about changing "dev->irq" when MSI-X is enabled, and it's only incidental that sysfs reflects that.

So we need to know the effect of changing dev->irq. Drivers may use the value of dev->irq, and I'm *guessing* this change shouldn't break them since we already do this for MSI, but I'd like some more expert opinion than mine :)

For MSI we have:

  msi_capability_init
    msi_setup_entry
      entry = alloc_msi_entry(nvec)
      entry->msi_attrib.default_irq = dev->irq;     /* Save IOAPIC IRQ */
    dev->irq = entry->irq;

  pci_msi_shutdown
    /* Restore dev->irq to its default pin-assertion IRQ */
    dev->irq = desc->msi_attrib.default_irq;

and for MSI-X we have:

msix_capability_init

    msix_setup_entries
      for (i = 0; i < nvec; i++)
        entry = alloc_msi_entry(1)
	entry->msi_attrib.default_irq = dev->irq;

pci_msix_shutdown
for_each_pci_msi_entry(entry, dev)
__pci_msix_desc_mask_irq
+ dev->irq = entry->msi_attrib.default_irq; # added by this patch

Things that seem strange to me:

- The msi_setup_entry() comment "Save IOAPIC IRQ" seems needlessly specific; maybe it should be "INTx IRQ".

- The pci_msi_shutdown() comment "Restore ... pin-assertion IRQ" should match the msi_setup_entry() one, e.g., maybe it should also be "INTx IRQ". There are no INTx or IOAPIC pins in PCIe.

- The only use of .default_irq is to save and restore dev->irq, so it looks like a per-device thing, not a per-vector thing.

In msi_setup_entry() there's only one msi_entry, so there's only one saved .default_irq.

In msix_setup_entries(), we get nvecs msi_entry structs, and we get a saved .default_irq in each one?

- In pci_msix_shutdown() we restore dev->irq from the .default_irq of whatever "entry" contains after the for_each_pci_msi_entry() loop. Is "entry" defined there? I don't know what the cursor contains after a list_for_each_entry() loop exits.

Quoted 14 lines
drivers/pci/msi.c +1 -0
 	return 0;
 
 out_avail:
@@ -1024,6 +1029,7 @@ static void pci_msix_shutdown(struct pci_dev *dev)
 	pci_msix_clear_and_set_ctrl(dev, PCI_MSIX_FLAGS_ENABLE, 0);
 	pci_intx_for_msi(dev, 1);
 	dev->msix_enabled = 0;
+	dev->irq = entry->msi_attrib.default_irq;
 	pcibios_alloc_irq(dev);
 }
 
Signature · 1 line
1.8.3.1
#3 lore raw

Re: [PATCH v2 1/2] PCI/MSI: Fix the confusing IRQ sysfs ABI for MSI-X

Hi Bjorn,

Quoted 90 lines · Bjorn Helgaas

[+cc Thomas, Marc]

Quoted 4 lines · Barry Song

From: Barry Song <[email protected]>

/sys/bus/pci/devices/.../irq sysfs ABI is very confusing at this moment especially for MSI-X cases.

AFAICT this patch *only* affects MSI-X. So are you saying the sysfs ABI is fine for MSI but confusing for MSI-X?

Quoted 27 lines

While MSI sets IRQ to the first
number in the vector, MSI-X does nothing for this though it saves default_irq in msix_setup_entries(). Weird the saved default_irq for MSI-X is never used in pci_msix_shutdown(), which is quite different with pci_msi_shutdown(). Thus, this patch moves to show the first IRQ number which is from the first msi_entry for MSI-X. Hopefully, this can make IRQ ABI more clear and more consistent.

---
 drivers/pci/msi.c | 6 ++++++
 1 file changed, 6 insertions(+)
drivers/pci/msi.c +1 -0
index 9232255..6bbf81b 100644
@@ -771,6 +771,7 @@ static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries,
 	int ret;
 	u16 control;
 	void __iomem *base;
+	struct msi_desc *desc;
 
 	/* Ensure MSI-X is disabled while it is set up */
 	pci_msix_clear_and_set_ctrl(dev, PCI_MSIX_FLAGS_ENABLE, 0);

drivers/pci/msi.c static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries, · line 820

Quoted 6 lines
drivers/pci/msi.c +3 -0
@@ -814,6 +815,10 @@ static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries,
 	pci_msix_clear_and_set_ctrl(dev, PCI_MSIX_FLAGS_MASKALL, 0);
 
 	pcibios_free_irq(dev);
+
+	desc = first_pci_msi_entry(dev);
+	dev->irq = desc->irq;

This change is not primarily about sysfs. This is about changing "dev->irq" when MSI-X is enabled, and it's only incidental that sysfs reflects that.

So we need to know the effect of changing dev->irq. Drivers may use the value of dev->irq, and I'm *guessing* this change shouldn't break them since we already do this for MSI, but I'd like some more expert opinion than mine :)

For MSI we have:

  msi_capability_init
    msi_setup_entry
      entry = alloc_msi_entry(nvec)
      entry->msi_attrib.default_irq = dev->irq;     /* Save IOAPIC IRQ */
    dev->irq = entry->irq;

  pci_msi_shutdown
    /* Restore dev->irq to its default pin-assertion IRQ */
    dev->irq = desc->msi_attrib.default_irq;

and for MSI-X we have:

msix_capability_init

    msix_setup_entries
      for (i = 0; i < nvec; i++)
        entry = alloc_msi_entry(1)
	entry->msi_attrib.default_irq = dev->irq;

pci_msix_shutdown
for_each_pci_msi_entry(entry, dev)
__pci_msix_desc_mask_irq
+ dev->irq = entry->msi_attrib.default_irq; # added by this patch

Things that seem strange to me:

- The msi_setup_entry() comment "Save IOAPIC IRQ" seems needlessly specific; maybe it should be "INTx IRQ".

- The pci_msi_shutdown() comment "Restore ... pin-assertion IRQ" should match the msi_setup_entry() one, e.g., maybe it should also be "INTx IRQ". There are no INTx or IOAPIC pins in PCIe.

Quoted 8 lines

- The only use of .default_irq is to save and restore dev->irq, so it looks like a per-device thing, not a per-vector thing.

In msi_setup_entry() there's only one msi_entry, so there's only one saved .default_irq.

In msix_setup_entries(), we get nvecs msi_entry structs, and we get a saved .default_irq in each one?

That's a key point.

Old-school PCI/MSI is represented by a single interrupt, and you *could* somehow make it relatively easy for drivers that only understand INTx to migrate to MSI if you replaced whatever is held in dev->irq (which should only represent the INTx mapping) with the MSI interrupt number. Which I guess is what the MSI code is doing.

This is the 21st century, and nobody should ever rely on such horror, but I'm sure we do have such drivers in the tree. Boo.

However, this *cannot* hold true for Multi-MSI, nor MSI-X, because there is a plurality of interrupts. Even worse, for MSI-X, there is zero guarantee that the allocated interrupts will be in a contiguous space.

Given that, what is dev->irq good for? "Absolutely Nothing! (say it again!)".

MSI-X is not something you can "accidentally" use. You have to actively embrace it. In all honesty, this patch tries to move in the wrong direction. If anything, we should kill this hack altogether and fix the (handful of?) drivers that rely on it. That'd actually be a good way to find whether they are still worth keeping in the tree. And if it breaks too many of them, then at least we'll know where we stand.

I'd be tempted to leave the below patch simmer in -next for a few weeks and see if how many people shout:

drivers/pci/msi.c +0 -5
index e5e75331b415..2be9a01cbe72 100644
@@ -591,7 +591,6 @@ msi_setup_entry(struct pci_dev *dev, int nvec, struct irq_affinity *affd)
 	entry->msi_attrib.is_virtual    = 0;
 	entry->msi_attrib.entry_nr	= 0;
 	entry->msi_attrib.maskbit	= !!(control & PCI_MSI_FLAGS_MASKBIT);
-	entry->msi_attrib.default_irq	= dev->irq;	/* Save IOAPIC IRQ */
 	entry->msi_attrib.multi_cap	= (control & PCI_MSI_FLAGS_QMASK) >> 1;
 	entry->msi_attrib.multiple	= ilog2(__roundup_pow_of_two(nvec));
 
@@ -682,7 +681,6 @@ static int msi_capability_init(struct pci_dev *dev, int nvec,
 	dev->msi_enabled = 1;
 
 	pcibios_free_irq(dev);
-	dev->irq = entry->irq;
 	return 0;
 }
 
@@ -742,7 +740,6 @@ static int msix_setup_entries(struct pci_dev *dev, void __iomem *base,
 		entry->msi_attrib.is_virtual =
 			entry->msi_attrib.entry_nr >= vec_count;
 
-		entry->msi_attrib.default_irq	= dev->irq;
 		entry->mask_base		= base;
 
 		addr = pci_msix_desc_addr(entry);
@@ -964,8 +961,6 @@ static void pci_msi_shutdown(struct pci_dev *dev)
 	mask = msi_mask(desc->msi_attrib.multi_cap);
 	msi_mask_irq(desc, mask, 0);
 
-	/* Restore dev->irq to its default pin-assertion IRQ */
-	dev->irq = desc->msi_attrib.default_irq;
 	pcibios_alloc_irq(dev);
 }
 
include/linux/msi.h +0 -2
index e8bdcb83172b..a631664c1c38 100644
@@ -114,7 +114,6 @@ struct ti_sci_inta_msi_desc {
  * @maskbit:	[PCI MSI/X] Mask-Pending bit supported?
  * @is_64:	[PCI MSI/X] Address size: 0=32bit 1=64bit
  * @entry_nr:	[PCI MSI/X] Entry which is described by this descriptor
- * @default_irq:[PCI MSI/X] The default pre-assigned non-MSI irq
  * @mask_pos:	[PCI MSI]   Mask register position
  * @mask_base:	[PCI MSI-X] Mask register base address
  * @platform:	[platform]  Platform device specific msi descriptor data
@@ -148,7 +147,6 @@ struct msi_desc {
 				u8	is_64		: 1;
 				u8	is_virtual	: 1;
 				u16	entry_nr;
-				unsigned default_irq;
 			} msi_attrib;
 			union {
 				u8	mask_pos;

Thanks,

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
author
#4 lore raw

Re: [PATCH v2 1/2] PCI/MSI: Fix the confusing IRQ sysfs ABI for MSI-X

Quoted 115 lines · Marc Zyngier

Hi Bjorn,

Quoted 90 lines · Bjorn Helgaas

[+cc Thomas, Marc]

Quoted 4 lines · Barry Song

From: Barry Song <[email protected]>

/sys/bus/pci/devices/.../irq sysfs ABI is very confusing at this moment especially for MSI-X cases.

AFAICT this patch *only* affects MSI-X. So are you saying the sysfs ABI is fine for MSI but confusing for MSI-X?

Quoted 27 lines

While MSI sets IRQ to the first
number in the vector, MSI-X does nothing for this though it saves default_irq in msix_setup_entries(). Weird the saved default_irq for MSI-X is never used in pci_msix_shutdown(), which is quite different with pci_msi_shutdown(). Thus, this patch moves to show the first IRQ number which is from the first msi_entry for MSI-X. Hopefully, this can make IRQ ABI more clear and more consistent.

---
 drivers/pci/msi.c | 6 ++++++
 1 file changed, 6 insertions(+)
drivers/pci/msi.c +1 -0
index 9232255..6bbf81b 100644
@@ -771,6 +771,7 @@ static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries,
    int ret;
    u16 control;
    void __iomem *base;
+   struct msi_desc *desc;
 
    /* Ensure MSI-X is disabled while it is set up */
    pci_msix_clear_and_set_ctrl(dev, PCI_MSIX_FLAGS_ENABLE, 0);

drivers/pci/msi.c static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries, · line 820

Quoted 6 lines
drivers/pci/msi.c +3 -0
@@ -814,6 +815,10 @@ static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries,
    pci_msix_clear_and_set_ctrl(dev, PCI_MSIX_FLAGS_MASKALL, 0);
 
    pcibios_free_irq(dev);
+
+   desc = first_pci_msi_entry(dev);
+   dev->irq = desc->irq;

This change is not primarily about sysfs. This is about changing "dev->irq" when MSI-X is enabled, and it's only incidental that sysfs reflects that.

So we need to know the effect of changing dev->irq. Drivers may use the value of dev->irq, and I'm *guessing* this change shouldn't break them since we already do this for MSI, but I'd like some more expert opinion than mine :)

For MSI we have:

  msi_capability_init
    msi_setup_entry
      entry = alloc_msi_entry(nvec)
      entry->msi_attrib.default_irq = dev->irq;     /* Save IOAPIC IRQ */
    dev->irq = entry->irq;

  pci_msi_shutdown
    /* Restore dev->irq to its default pin-assertion IRQ */
    dev->irq = desc->msi_attrib.default_irq;

and for MSI-X we have:

  msix_capability_init
    msix_setup_entries
      for (i = 0; i < nvec; i++)
        entry = alloc_msi_entry(1)
      entry->msi_attrib.default_irq = dev->irq;

pci_msix_shutdown
for_each_pci_msi_entry(entry, dev)
__pci_msix_desc_mask_irq
+ dev->irq = entry->msi_attrib.default_irq; # added by this patch

Things that seem strange to me:

- The msi_setup_entry() comment "Save IOAPIC IRQ" seems needlessly specific; maybe it should be "INTx IRQ".

- The pci_msi_shutdown() comment "Restore ... pin-assertion IRQ" should match the msi_setup_entry() one, e.g., maybe it should also be "INTx IRQ". There are no INTx or IOAPIC pins in PCIe.

Quoted 8 lines

- The only use of .default_irq is to save and restore dev->irq, so it looks like a per-device thing, not a per-vector thing.

In msi_setup_entry() there's only one msi_entry, so there's only one saved .default_irq.

In msix_setup_entries(), we get nvecs msi_entry structs, and we get a saved .default_irq in each one?

That's a key point.

Old-school PCI/MSI is represented by a single interrupt, and you *could* somehow make it relatively easy for drivers that only understand INTx to migrate to MSI if you replaced whatever is held in dev->irq (which should only represent the INTx mapping) with the MSI interrupt number. Which I guess is what the MSI code is doing.

This is the 21st century, and nobody should ever rely on such horror, but I'm sure we do have such drivers in the tree. Boo.

Quoted 8 lines

However, this *cannot* hold true for Multi-MSI, nor MSI-X, because there is a plurality of interrupts. Even worse, for MSI-X, there is zero guarantee that the allocated interrupts will be in a contiguous space.

Given that, what is dev->irq good for? "Absolutely Nothing! (say it again!)".

The only thing is that dev->irq is an sysfs ABI to userspace. Due to the inconsistency
between legacy PCI INTx, MSI, MSI-X, this ABI should have been absolutely broken nowadays.
This is actually what the patchset was originally aiming at to fix.

One more question from me is that does dev->irq actually hold any valid hardware INTx
information while hardware is using MSI-X? At least in my hardware, sysfs ABI for PCI is all "0".

root@ubuntu:/sys/devices/pci0000:7c/0000:7c:00.0/0000:7d:00.3# cat irq
0
root@ubuntu:/sys/devices/pci0000:7c/0000:7c:00.0/0000:7d:00.3# ls -l msi_irqs/*
-r--r--r-- 1 root root 4096 Aug 21 22:04 msi_irqs/499
-r--r--r-- 1 root root 4096 Aug 21 22:04 msi_irqs/500
-r--r--r-- 1 root root 4096 Aug 21 22:04 msi_irqs/501
...
root@ubuntu:/sys/devices/pci0000:7c/0000:7c:00.0/0000:7d:00.3# cat msi_irqs/499
msix

Not quite sure how it is going on different hardware platforms.

Quoted 10 lines

MSI-X is not something you can "accidentally" use. You have to actively embrace it. In all honesty, this patch tries to move in the wrong direction. If anything, we should kill this hack altogether and fix the (handful of?) drivers that rely on it. That'd actually be a good way to find whether they are still worth keeping in the tree. And if it breaks too many of them, then at least we'll know where we stand.

I'd be tempted to leave the below patch simmer in -next for a few weeks and see if how many people shout:

This looks like a more proper direction to go.
but here i am wondering how sysfs ABI document should follow the below change doc is patch 2/2:
https://lore.kernel.org/lkml/[email protected]/

On the other hand, my feeling is that nobody should depend on sysfs irq entry nowadays.
For example, userspace irqbalance is actually using /sys/devices/.../msi_irqs/ So probably we should set this ABI invisible when devices are using MSI or MSI-X?

Quoted 59 lines
drivers/pci/msi.c +0 -5
index e5e75331b415..2be9a01cbe72 100644
@@ -591,7 +591,6 @@ msi_setup_entry(struct pci_dev *dev, int nvec, struct irq_affinity *affd)
        entry->msi_attrib.is_virtual    = 0;
        entry->msi_attrib.entry_nr      = 0;
        entry->msi_attrib.maskbit       = !!(control & PCI_MSI_FLAGS_MASKBIT);
-       entry->msi_attrib.default_irq   = dev->irq;     /* Save IOAPIC IRQ */
        entry->msi_attrib.multi_cap     = (control & PCI_MSI_FLAGS_QMASK) >> 1;
        entry->msi_attrib.multiple      = ilog2(__roundup_pow_of_two(nvec));
 
@@ -682,7 +681,6 @@ static int msi_capability_init(struct pci_dev *dev, int nvec,
        dev->msi_enabled = 1;
 
        pcibios_free_irq(dev);
-       dev->irq = entry->irq;
        return 0;
 }
 
@@ -742,7 +740,6 @@ static int msix_setup_entries(struct pci_dev *dev, void __iomem *base,
                entry->msi_attrib.is_virtual =
                        entry->msi_attrib.entry_nr >= vec_count;
 
-               entry->msi_attrib.default_irq   = dev->irq;
                entry->mask_base                = base;
 
                addr = pci_msix_desc_addr(entry);
@@ -964,8 +961,6 @@ static void pci_msi_shutdown(struct pci_dev *dev)
        mask = msi_mask(desc->msi_attrib.multi_cap);
        msi_mask_irq(desc, mask, 0);
 
-       /* Restore dev->irq to its default pin-assertion IRQ */
-       dev->irq = desc->msi_attrib.default_irq;
        pcibios_alloc_irq(dev);
 }
 
include/linux/msi.h +0 -2
index e8bdcb83172b..a631664c1c38 100644
@@ -114,7 +114,6 @@ struct ti_sci_inta_msi_desc {
  * @maskbit:   [PCI MSI/X] Mask-Pending bit supported?
  * @is_64:     [PCI MSI/X] Address size: 0=32bit 1=64bit
  * @entry_nr:  [PCI MSI/X] Entry which is described by this descriptor
- * @default_irq:[PCI MSI/X] The default pre-assigned non-MSI irq
  * @mask_pos:  [PCI MSI]   Mask register position
  * @mask_base: [PCI MSI-X] Mask register base address
  * @platform:  [platform]  Platform device specific msi descriptor data
@@ -148,7 +147,6 @@ struct msi_desc {
                                u8      is_64           : 1;
                                u8      is_virtual      : 1;
                                u16     entry_nr;
-                               unsigned default_irq;
                        } msi_attrib;
                        union {
                                u8      mask_pos;
Quoted 6 lines

Thanks,

        M.

--
Without deviation from the norm, progress is not possible.

Thanks
barry

author
#5 lore raw

Re: [PATCH v2 1/2] PCI/MSI: Fix the confusing IRQ sysfs ABI for MSI-X

Quoted 166 lines · Barry Song
Quoted 115 lines · Marc Zyngier

Hi Bjorn,

Quoted 90 lines · Bjorn Helgaas

[+cc Thomas, Marc]

Quoted 4 lines · Barry Song

From: Barry Song <[email protected]>

/sys/bus/pci/devices/.../irq sysfs ABI is very confusing at this moment especially for MSI-X cases.

AFAICT this patch *only* affects MSI-X. So are you saying the sysfs ABI is fine for MSI but confusing for MSI-X?

Quoted 27 lines

While MSI sets IRQ to the first
number in the vector, MSI-X does nothing for this though it saves default_irq in msix_setup_entries(). Weird the saved default_irq for MSI-X is never used in pci_msix_shutdown(), which is quite different with pci_msi_shutdown(). Thus, this patch moves to show the first IRQ number which is from the first msi_entry for MSI-X. Hopefully, this can make IRQ ABI more clear and more consistent.

---
 drivers/pci/msi.c | 6 ++++++
 1 file changed, 6 insertions(+)
drivers/pci/msi.c +1 -0
index 9232255..6bbf81b 100644
@@ -771,6 +771,7 @@ static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries,
    int ret;
    u16 control;
    void __iomem *base;
+   struct msi_desc *desc;
 
    /* Ensure MSI-X is disabled while it is set up */
    pci_msix_clear_and_set_ctrl(dev, PCI_MSIX_FLAGS_ENABLE, 0);

drivers/pci/msi.c static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries, · line 820

Quoted 6 lines
drivers/pci/msi.c +3 -0
@@ -814,6 +815,10 @@ static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries,
    pci_msix_clear_and_set_ctrl(dev, PCI_MSIX_FLAGS_MASKALL, 0);
 
    pcibios_free_irq(dev);
+
+   desc = first_pci_msi_entry(dev);
+   dev->irq = desc->irq;

This change is not primarily about sysfs. This is about changing "dev->irq" when MSI-X is enabled, and it's only incidental that sysfs reflects that.

So we need to know the effect of changing dev->irq. Drivers may use the value of dev->irq, and I'm *guessing* this change shouldn't break them since we already do this for MSI, but I'd like some more expert opinion than mine :)

For MSI we have:

  msi_capability_init
    msi_setup_entry
      entry = alloc_msi_entry(nvec)
      entry->msi_attrib.default_irq = dev->irq;     /* Save IOAPIC IRQ */
    dev->irq = entry->irq;

  pci_msi_shutdown
    /* Restore dev->irq to its default pin-assertion IRQ */
    dev->irq = desc->msi_attrib.default_irq;

and for MSI-X we have:

  msix_capability_init
    msix_setup_entries
      for (i = 0; i < nvec; i++)
        entry = alloc_msi_entry(1)
      entry->msi_attrib.default_irq = dev->irq;

pci_msix_shutdown
for_each_pci_msi_entry(entry, dev)
__pci_msix_desc_mask_irq
+ dev->irq = entry->msi_attrib.default_irq; # added by this patch

Things that seem strange to me:

- The msi_setup_entry() comment "Save IOAPIC IRQ" seems needlessly specific; maybe it should be "INTx IRQ".

- The pci_msi_shutdown() comment "Restore ... pin-assertion IRQ" should match the msi_setup_entry() one, e.g., maybe it should also be "INTx IRQ". There are no INTx or IOAPIC pins in PCIe.

Quoted 8 lines

- The only use of .default_irq is to save and restore dev->irq, so it looks like a per-device thing, not a per-vector thing.

In msi_setup_entry() there's only one msi_entry, so there's only one saved .default_irq.

In msix_setup_entries(), we get nvecs msi_entry structs, and we get a saved .default_irq in each one?

That's a key point.

Old-school PCI/MSI is represented by a single interrupt, and you *could* somehow make it relatively easy for drivers that only understand INTx to migrate to MSI if you replaced whatever is held in dev->irq (which should only represent the INTx mapping) with the MSI interrupt number. Which I guess is what the MSI code is doing.

This is the 21st century, and nobody should ever rely on such horror, but I'm sure we do have such drivers in the tree. Boo.

Quoted 8 lines

However, this *cannot* hold true for Multi-MSI, nor MSI-X, because there is a plurality of interrupts. Even worse, for MSI-X, there is zero guarantee that the allocated interrupts will be in a contiguous space.

Given that, what is dev->irq good for? "Absolutely Nothing! (say it again!)".

The only thing is that dev->irq is an sysfs ABI to userspace. Due to the inconsistency
between legacy PCI INTx, MSI, MSI-X, this ABI should have been absolutely broken nowadays.
This is actually what the patchset was originally aiming at to fix.

One more question from me is that does dev->irq actually hold any valid hardware INTx
information while hardware is using MSI-X? At least in my hardware, sysfs ABI for PCI is all "0".

root@ubuntu:/sys/devices/pci0000:7c/0000:7c:00.0/0000:7d:00.3# cat irq
0
root@ubuntu:/sys/devices/pci0000:7c/0000:7c:00.0/0000:7d:00.3# ls -l msi_irqs/*
(quoted diff) +0 -3
-r--r--r-- 1 root root 4096 Aug 21 22:04 msi_irqs/499
-r--r--r-- 1 root root 4096 Aug 21 22:04 msi_irqs/500
-r--r--r-- 1 root root 4096 Aug 21 22:04 msi_irqs/501
...
root@ubuntu:/sys/devices/pci0000:7c/0000:7c:00.0/0000:7d:00.3# cat msi_irqs/499
msix

Not quite sure how it is going on different hardware platforms.

Quoted 10 lines

MSI-X is not something you can "accidentally" use. You have to actively embrace it. In all honesty, this patch tries to move in the wrong direction. If anything, we should kill this hack altogether and fix the (handful of?) drivers that rely on it. That'd actually be a good way to find whether they are still worth keeping in the tree. And if it breaks too many of them, then at least we'll know where we stand.

I'd be tempted to leave the below patch simmer in -next for a few weeks and see if how many people shout:

This looks like a more proper direction to go.
but here i am wondering how sysfs ABI document should follow the below change doc is patch 2/2:
https://lore.kernel.org/lkml/[email protected]/

Quoted 5 lines

On the other hand, my feeling is that nobody should depend on sysfs irq entry nowadays.
For example, userspace irqbalance is actually using /sys/devices/.../msi_irqs/ So probably we should set this ABI invisible when devices are using MSI or MSI-X?

i mean something like the below,

drivers/pci/pci-sysfs.c +1 -0
index 5d63df7..1323841 100644
@@ -26,6 +26,7 @@
 #include <linux/slab.h>
 #include <linux/vgaarb.h>
 #include <linux/pm_runtime.h>
+#include <linux/msi.h>
 #include <linux/of.h>
 #include "pci.h"
 
@@ -1437,6 +1438,16 @@ static umode_t pci_dev_attrs_are_visible(struct

kobject *kobj,

                if ((pdev->class >> 8) != PCI_CLASS_DISPLAY_VGA)
                        return 0;

+#ifdef CONFIG_PCI_MSI
+ /*

+        * if devices are MSI and MSI-X, IRQ sysfs ABI is meaningless
+        * and broken
+        */

+ if (a == &dev_attr_irq.attr)
+ if (first_pci_msi_entry(pdev))
+ return 0;
+#endif
+

        return a->mode;

}

Quoted 60 lines
Quoted 60 lines
drivers/pci/msi.c +0 -5
index e5e75331b415..2be9a01cbe72 100644
@@ -591,7 +591,6 @@ msi_setup_entry(struct pci_dev *dev, int nvec, struct irq_affinity *affd)
        entry->msi_attrib.is_virtual    = 0;
        entry->msi_attrib.entry_nr      = 0;
        entry->msi_attrib.maskbit       = !!(control & PCI_MSI_FLAGS_MASKBIT);
-       entry->msi_attrib.default_irq   = dev->irq;     /* Save IOAPIC IRQ */
        entry->msi_attrib.multi_cap     = (control & PCI_MSI_FLAGS_QMASK) >> 1;
        entry->msi_attrib.multiple      = ilog2(__roundup_pow_of_two(nvec));
 
@@ -682,7 +681,6 @@ static int msi_capability_init(struct pci_dev *dev, int nvec,
        dev->msi_enabled = 1;
 
        pcibios_free_irq(dev);
-       dev->irq = entry->irq;
        return 0;
 }
 
@@ -742,7 +740,6 @@ static int msix_setup_entries(struct pci_dev *dev, void __iomem *base,
                entry->msi_attrib.is_virtual =
                        entry->msi_attrib.entry_nr >= vec_count;
 
-               entry->msi_attrib.default_irq   = dev->irq;
                entry->mask_base                = base;
 
                addr = pci_msix_desc_addr(entry);
@@ -964,8 +961,6 @@ static void pci_msi_shutdown(struct pci_dev *dev)
        mask = msi_mask(desc->msi_attrib.multi_cap);
        msi_mask_irq(desc, mask, 0);
 
-       /* Restore dev->irq to its default pin-assertion IRQ */
-       dev->irq = desc->msi_attrib.default_irq;
        pcibios_alloc_irq(dev);
 }
 
include/linux/msi.h +0 -2
index e8bdcb83172b..a631664c1c38 100644
@@ -114,7 +114,6 @@ struct ti_sci_inta_msi_desc {
  * @maskbit:   [PCI MSI/X] Mask-Pending bit supported?
  * @is_64:     [PCI MSI/X] Address size: 0=32bit 1=64bit
  * @entry_nr:  [PCI MSI/X] Entry which is described by this descriptor
- * @default_irq:[PCI MSI/X] The default pre-assigned non-MSI irq
  * @mask_pos:  [PCI MSI]   Mask register position
  * @mask_base: [PCI MSI-X] Mask register base address
  * @platform:  [platform]  Platform device specific msi descriptor data
@@ -148,7 +147,6 @@ struct msi_desc {
                                u8      is_64           : 1;
                                u8      is_virtual      : 1;
                                u16     entry_nr;
-                               unsigned default_irq;
                        } msi_attrib;
                        union {
                                u8      mask_pos;
Quoted 7 lines
Quoted 6 lines

Thanks,

        M.

--
Without deviation from the norm, progress is not possible.

Thanks
barry

#6 lore raw

Re: [PATCH v2 1/2] PCI/MSI: Fix the confusing IRQ sysfs ABI for MSI-X

On Sat, 21 Aug 2021 23:41:17 +0100,
Barry Song <[email protected]> wrote:

[...]

Quoted 24 lines
Quoted 2 lines

So probably we should set this ABI invisible when devices are using MSI or MSI-X?

i mean something like the below,

drivers/pci/pci-sysfs.c +3 -0
index 5d63df7..1323841 100644
@@ -26,6 +26,7 @@
 #include <linux/slab.h>
 #include <linux/vgaarb.h>
 #include <linux/pm_runtime.h>
+#include <linux/msi.h>
 #include <linux/of.h>
 #include "pci.h"
 
@@ -1437,6 +1438,16 @@ static umode_t pci_dev_attrs_are_visible(struct
 kobject *kobj,
                if ((pdev->class >> 8) != PCI_CLASS_DISPLAY_VGA)
                        return 0;
 
+#ifdef CONFIG_PCI_MSI
+       /*

drivers/pci/pci-sysfs.c static umode_t pci_dev_attrs_are_visible(struct · line 1453

Quoted 10 lines
drivers/pci/pci-sysfs.c +8 -0
+        * if devices are MSI and MSI-X, IRQ sysfs ABI is meaningless
+        * and broken
+        */
+       if (a == &dev_attr_irq.attr)
+               if (first_pci_msi_entry(pdev))
+                       return 0;
+#endif
+
        return a->mode;
 }

I don't think you can break what we have today. Whatever change we make to the kernel internals, the userspace view must stay unchanged.

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
#7 lore raw

Re: [PATCH v2 1/2] PCI/MSI: Fix the confusing IRQ sysfs ABI for MSI-X

In reply to Barry Song
Quoted 194 lines · Barry Song
Quoted 166 lines · Barry Song
Quoted 115 lines · Marc Zyngier

Hi Bjorn,

Quoted 90 lines · Bjorn Helgaas

[+cc Thomas, Marc]

Quoted 4 lines · Barry Song

From: Barry Song <[email protected]>

/sys/bus/pci/devices/.../irq sysfs ABI is very confusing at this moment especially for MSI-X cases.

AFAICT this patch *only* affects MSI-X. So are you saying the sysfs ABI is fine for MSI but confusing for MSI-X?

Quoted 27 lines

While MSI sets IRQ to the first
number in the vector, MSI-X does nothing for this though it saves default_irq in msix_setup_entries(). Weird the saved default_irq for MSI-X is never used in pci_msix_shutdown(), which is quite different with pci_msi_shutdown(). Thus, this patch moves to show the first IRQ number which is from the first msi_entry for MSI-X. Hopefully, this can make IRQ ABI more clear and more consistent.

---
 drivers/pci/msi.c | 6 ++++++
 1 file changed, 6 insertions(+)
drivers/pci/msi.c +1 -0
index 9232255..6bbf81b 100644
@@ -771,6 +771,7 @@ static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries,
    int ret;
    u16 control;
    void __iomem *base;
+   struct msi_desc *desc;
 
    /* Ensure MSI-X is disabled while it is set up */
    pci_msix_clear_and_set_ctrl(dev, PCI_MSIX_FLAGS_ENABLE, 0);

drivers/pci/msi.c static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries, · line 820

Quoted 6 lines
drivers/pci/msi.c +3 -0
@@ -814,6 +815,10 @@ static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries,
    pci_msix_clear_and_set_ctrl(dev, PCI_MSIX_FLAGS_MASKALL, 0);
 
    pcibios_free_irq(dev);
+
+   desc = first_pci_msi_entry(dev);
+   dev->irq = desc->irq;

This change is not primarily about sysfs. This is about changing "dev->irq" when MSI-X is enabled, and it's only incidental that sysfs reflects that.

So we need to know the effect of changing dev->irq. Drivers may use the value of dev->irq, and I'm *guessing* this change shouldn't break them since we already do this for MSI, but I'd like some more expert opinion than mine :)

For MSI we have:

  msi_capability_init
    msi_setup_entry
      entry = alloc_msi_entry(nvec)
      entry->msi_attrib.default_irq = dev->irq;     /* Save IOAPIC IRQ */
    dev->irq = entry->irq;

  pci_msi_shutdown
    /* Restore dev->irq to its default pin-assertion IRQ */
    dev->irq = desc->msi_attrib.default_irq;

and for MSI-X we have:

  msix_capability_init
    msix_setup_entries
      for (i = 0; i < nvec; i++)
        entry = alloc_msi_entry(1)
      entry->msi_attrib.default_irq = dev->irq;

pci_msix_shutdown
for_each_pci_msi_entry(entry, dev)
__pci_msix_desc_mask_irq
+ dev->irq = entry->msi_attrib.default_irq; # added by this patch

Things that seem strange to me:

- The msi_setup_entry() comment "Save IOAPIC IRQ" seems needlessly specific; maybe it should be "INTx IRQ".

- The pci_msi_shutdown() comment "Restore ... pin-assertion IRQ" should match the msi_setup_entry() one, e.g., maybe it should also be "INTx IRQ". There are no INTx or IOAPIC pins in PCIe.

Quoted 8 lines

- The only use of .default_irq is to save and restore dev->irq, so it looks like a per-device thing, not a per-vector thing.

In msi_setup_entry() there's only one msi_entry, so there's only one saved .default_irq.

In msix_setup_entries(), we get nvecs msi_entry structs, and we get a saved .default_irq in each one?

That's a key point.

Old-school PCI/MSI is represented by a single interrupt, and you *could* somehow make it relatively easy for drivers that only understand INTx to migrate to MSI if you replaced whatever is held in dev->irq (which should only represent the INTx mapping) with the MSI interrupt number. Which I guess is what the MSI code is doing.

This is the 21st century, and nobody should ever rely on such horror, but I'm sure we do have such drivers in the tree. Boo.

Quoted 8 lines

However, this *cannot* hold true for Multi-MSI, nor MSI-X, because there is a plurality of interrupts. Even worse, for MSI-X, there is zero guarantee that the allocated interrupts will be in a contiguous space.

Given that, what is dev->irq good for? "Absolutely Nothing! (say it again!)".

The only thing is that dev->irq is an sysfs ABI to userspace. Due to the inconsistency
between legacy PCI INTx, MSI, MSI-X, this ABI should have been absolutely broken nowadays.
This is actually what the patchset was originally aiming at to fix.

One more question from me is that does dev->irq actually hold any valid hardware INTx
information while hardware is using MSI-X? At least in my hardware, sysfs ABI for PCI is all "0".

root@ubuntu:/sys/devices/pci0000:7c/0000:7c:00.0/0000:7d:00.3# cat irq
0
root@ubuntu:/sys/devices/pci0000:7c/0000:7c:00.0/0000:7d:00.3# ls -l msi_irqs/*
(quoted diff) +0 -3
-r--r--r-- 1 root root 4096 Aug 21 22:04 msi_irqs/499
-r--r--r-- 1 root root 4096 Aug 21 22:04 msi_irqs/500
-r--r--r-- 1 root root 4096 Aug 21 22:04 msi_irqs/501
...
root@ubuntu:/sys/devices/pci0000:7c/0000:7c:00.0/0000:7d:00.3# cat msi_irqs/499
msix

Not quite sure how it is going on different hardware platforms.

Quoted 10 lines

MSI-X is not something you can "accidentally" use. You have to actively embrace it. In all honesty, this patch tries to move in the wrong direction. If anything, we should kill this hack altogether and fix the (handful of?) drivers that rely on it. That'd actually be a good way to find whether they are still worth keeping in the tree. And if it breaks too many of them, then at least we'll know where we stand.

I'd be tempted to leave the below patch simmer in -next for a few weeks and see if how many people shout:

This looks like a more proper direction to go.
but here i am wondering how sysfs ABI document should follow the below change doc is patch 2/2:
https://lore.kernel.org/lkml/[email protected]/

Quoted 5 lines

On the other hand, my feeling is that nobody should depend on sysfs irq entry nowadays.
For example, userspace irqbalance is actually using /sys/devices/.../msi_irqs/ So probably we should set this ABI invisible when devices are using MSI or MSI-X?

i mean something like the below,

drivers/pci/pci-sysfs.c +3 -0
index 5d63df7..1323841 100644
@@ -26,6 +26,7 @@
 #include <linux/slab.h>
 #include <linux/vgaarb.h>
 #include <linux/pm_runtime.h>
+#include <linux/msi.h>
 #include <linux/of.h>
 #include "pci.h"
 
@@ -1437,6 +1438,16 @@ static umode_t pci_dev_attrs_are_visible(struct
 kobject *kobj,
                if ((pdev->class >> 8) != PCI_CLASS_DISPLAY_VGA)
                        return 0;
 
+#ifdef CONFIG_PCI_MSI
+       /*

drivers/pci/pci-sysfs.c static umode_t pci_dev_attrs_are_visible(struct · line 1453

Quoted 10 lines
drivers/pci/pci-sysfs.c +8 -0
+        * if devices are MSI and MSI-X, IRQ sysfs ABI is meaningless
+        * and broken
+        */
+       if (a == &dev_attr_irq.attr)
+               if (first_pci_msi_entry(pdev))
+                       return 0;
+#endif
+
        return a->mode;
 }

I think this idea has been discarded anyway, but sysfs doesn't work this way. The .is_visible() function is evaluated once at device_add()-time, i.e., during enumeration, so there's no way to dynamically change the visibility as the driver enables/disables MSI.

I *wish* sysfs had more flexibility like this, though.

#8 lore raw

Re: [PATCH v2 1/2] PCI/MSI: Fix the confusing IRQ sysfs ABI for MSI-X

In reply to Barry Song
Quoted 8 lines · Barry Song
Quoted 6 lines · Marc Zyngier

Hi Bjorn,

Quoted 1 line · Bjorn Helgaas

[...]

Quoted 19 lines
Quoted 19 lines
Quoted 2 lines

In msix_setup_entries(), we get nvecs msi_entry structs, and we get a saved .default_irq in each one?

That's a key point.

Old-school PCI/MSI is represented by a single interrupt, and you *could* somehow make it relatively easy for drivers that only understand INTx to migrate to MSI if you replaced whatever is held in dev->irq (which should only represent the INTx mapping) with the MSI interrupt number. Which I guess is what the MSI code is doing.

This is the 21st century, and nobody should ever rely on such horror, but I'm sure we do have such drivers in the tree. Boo.

However, this *cannot* hold true for Multi-MSI, nor MSI-X, because there is a plurality of interrupts. Even worse, for MSI-X, there is zero guarantee that the allocated interrupts will be in a contiguous space.

Quoted 8 lines
Quoted 3 lines

Given that, what is dev->irq good for? "Absolutely Nothing! (say it again!)".

The only thing is that dev->irq is an sysfs ABI to userspace. Due to the inconsistency between legacy PCI INTx, MSI, MSI-X, this ABI should have been absolutely broken nowadays. This is actually what the patchset was originally aiming at to fix.

I do not think we should expose more of a broken abstraction to userspace. We will have to carry on exposing the first MSI in this field forever, but it doesn't mean we should have to do it for MSI-X.

Quoted 3 lines

One more question from me is that does dev->irq actually hold any valid hardware INTx information while hardware is using MSI-X? At least in my hardware, sysfs ABI for PCI is all "0".

That's probably because nothing actually configured the interrupt, or that there is no INTx implementation. I have that on systems with pretty dodgy (or incomplete) firmware.

Quoted 11 lines
root@ubuntu:/sys/devices/pci0000:7c/0000:7c:00.0/0000:7d:00.3# cat irq
0
root@ubuntu:/sys/devices/pci0000:7c/0000:7c:00.0/0000:7d:00.3# ls -l msi_irqs/*
(quoted diff) +0 -3
-r--r--r-- 1 root root 4096 Aug 21 22:04 msi_irqs/499
-r--r--r-- 1 root root 4096 Aug 21 22:04 msi_irqs/500
-r--r--r-- 1 root root 4096 Aug 21 22:04 msi_irqs/501
...
root@ubuntu:/sys/devices/pci0000:7c/0000:7c:00.0/0000:7d:00.3# cat msi_irqs/499
msix
Quoted 1 line

Not quite sure how it is going on different hardware platforms.

My D05 does that as well, and it doesn't expose any INTx support.

Quoted 12 lines
Quoted 11 lines

MSI-X is not something you can "accidentally" use. You have to actively embrace it. In all honesty, this patch tries to move in the wrong direction. If anything, we should kill this hack altogether and fix the (handful of?) drivers that rely on it. That'd actually be a good way to find whether they are still worth keeping in the tree. And if it breaks too many of them, then at least we'll know where we stand.

I'd be tempted to leave the below patch simmer in -next for a few weeks and see if how many people shout:

Quoted 7 lines

This looks like a more proper direction to go.
but here i am wondering how sysfs ABI document should follow the below change doc is patch 2/2:
https://lore.kernel.org/lkml/[email protected]/

On the other hand, my feeling is that nobody should depend on sysfs irq entry nowadays.

Too late. It is there, and we need to preserve it. I just don't think feeding it more erroneous information is the right thing to do.

My patch was only dealing with the kernel side of things, not the userspace ABI. That ABI should be carried on unchanged.

Quoted 3 lines

For example, userspace irqbalance is actually using /sys/devices/.../msi_irqs/ So probably we should set this ABI invisible when devices are using MSI or MSI-X?

Can it actually be made optional? I don't believe we can.

Thanks,

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
author
#9 lore raw

Re: [PATCH v2 1/2] PCI/MSI: Fix the confusing IRQ sysfs ABI for MSI-X

Quoted 86 lines · Marc Zyngier
Quoted 8 lines · Barry Song
Quoted 6 lines · Marc Zyngier

Hi Bjorn,

Quoted 1 line · Bjorn Helgaas

[...]

Quoted 19 lines
Quoted 19 lines
Quoted 2 lines

In msix_setup_entries(), we get nvecs msi_entry structs, and we get a saved .default_irq in each one?

That's a key point.

Old-school PCI/MSI is represented by a single interrupt, and you *could* somehow make it relatively easy for drivers that only understand INTx to migrate to MSI if you replaced whatever is held in dev->irq (which should only represent the INTx mapping) with the MSI interrupt number. Which I guess is what the MSI code is doing.

This is the 21st century, and nobody should ever rely on such horror, but I'm sure we do have such drivers in the tree. Boo.

However, this *cannot* hold true for Multi-MSI, nor MSI-X, because there is a plurality of interrupts. Even worse, for MSI-X, there is zero guarantee that the allocated interrupts will be in a contiguous space.

Quoted 8 lines
Quoted 3 lines

Given that, what is dev->irq good for? "Absolutely Nothing! (say it again!)".

The only thing is that dev->irq is an sysfs ABI to userspace. Due to the inconsistency between legacy PCI INTx, MSI, MSI-X, this ABI should have been absolutely broken nowadays. This is actually what the patchset was originally aiming at to fix.

I do not think we should expose more of a broken abstraction to userspace. We will have to carry on exposing the first MSI in this field forever, but it doesn't mean we should have to do it for MSI-X.

Quoted 3 lines

One more question from me is that does dev->irq actually hold any valid hardware INTx information while hardware is using MSI-X? At least in my hardware, sysfs ABI for PCI is all "0".

That's probably because nothing actually configured the interrupt, or that there is no INTx implementation. I have that on systems with pretty dodgy (or incomplete) firmware.

Quoted 11 lines
root@ubuntu:/sys/devices/pci0000:7c/0000:7c:00.0/0000:7d:00.3# cat irq
0
root@ubuntu:/sys/devices/pci0000:7c/0000:7c:00.0/0000:7d:00.3# ls -l msi_irqs/*
(quoted diff) +0 -3
-r--r--r-- 1 root root 4096 Aug 21 22:04 msi_irqs/499
-r--r--r-- 1 root root 4096 Aug 21 22:04 msi_irqs/500
-r--r--r-- 1 root root 4096 Aug 21 22:04 msi_irqs/501
...
root@ubuntu:/sys/devices/pci0000:7c/0000:7c:00.0/0000:7d:00.3# cat msi_irqs/499
msix
Quoted 1 line

Not quite sure how it is going on different hardware platforms.

My D05 does that as well, and it doesn't expose any INTx support.

Quoted 17 lines
Quoted 11 lines

MSI-X is not something you can "accidentally" use. You have to actively embrace it. In all honesty, this patch tries to move in the wrong direction. If anything, we should kill this hack altogether and fix the (handful of?) drivers that rely on it. That'd actually be a good way to find whether they are still worth keeping in the tree. And if it breaks too many of them, then at least we'll know where we stand.

I'd be tempted to leave the below patch simmer in -next for a few weeks and see if how many people shout:

This looks like a more proper direction to go.
but here i am wondering how sysfs ABI document should follow the below change doc is patch 2/2:
https://lore.kernel.org/lkml/[email protected]/

Quoted 8 lines
Quoted 2 lines

On the other hand, my feeling is that nobody should depend on sysfs irq entry nowadays.

Too late. It is there, and we need to preserve it. I just don't think feeding it more erroneous information is the right thing to do.

My patch was only dealing with the kernel side of things, not the userspace ABI. That ABI should be carried on unchanged.

it seems this isn't true. your patch is also changing userspace ABI as long as you change pci_dev->irq
which will be shown in sysfs irq entry.

if we don't want to change the behaviour of any existing ABI, it seems the only thing we can do here
to document it well in ABI doc. i actually doubt anyone has really understood what the irq entry
is really showing.

Quoted 6 lines
Quoted 5 lines

For example, userspace irqbalance is actually using /sys/devices/.../msi_irqs/ So probably we should set this ABI invisible when devices are using MSI or MSI-X?

Quoted 8 lines

Can it actually be made optional? I don't believe we can.

Thanks,

        M.

--
Without deviation from the norm, progress is not possible.

Thanks
barry

#10 lore raw

Re: [PATCH v2 1/2] PCI/MSI: Fix the confusing IRQ sysfs ABI for MSI-X

Quoted 94 lines · Barry Song
Quoted 92 lines · Marc Zyngier
Quoted 8 lines · Barry Song
Quoted 6 lines · Marc Zyngier

Hi Bjorn,

Quoted 1 line · Bjorn Helgaas

[...]

Quoted 19 lines
Quoted 19 lines
Quoted 2 lines

In msix_setup_entries(), we get nvecs msi_entry structs, and we get a saved .default_irq in each one?

That's a key point.

Old-school PCI/MSI is represented by a single interrupt, and you *could* somehow make it relatively easy for drivers that only understand INTx to migrate to MSI if you replaced whatever is held in dev->irq (which should only represent the INTx mapping) with the MSI interrupt number. Which I guess is what the MSI code is doing.

This is the 21st century, and nobody should ever rely on such horror, but I'm sure we do have such drivers in the tree. Boo.

However, this *cannot* hold true for Multi-MSI, nor MSI-X, because there is a plurality of interrupts. Even worse, for MSI-X, there is zero guarantee that the allocated interrupts will be in a contiguous space.

Quoted 8 lines
Quoted 3 lines

Given that, what is dev->irq good for? "Absolutely Nothing! (say it again!)".

The only thing is that dev->irq is an sysfs ABI to userspace. Due to the inconsistency between legacy PCI INTx, MSI, MSI-X, this ABI should have been absolutely broken nowadays. This is actually what the patchset was originally aiming at to fix.

I do not think we should expose more of a broken abstraction to userspace. We will have to carry on exposing the first MSI in this field forever, but it doesn't mean we should have to do it for MSI-X.

Quoted 3 lines

One more question from me is that does dev->irq actually hold any valid hardware INTx information while hardware is using MSI-X? At least in my hardware, sysfs ABI for PCI is all "0".

That's probably because nothing actually configured the interrupt, or that there is no INTx implementation. I have that on systems with pretty dodgy (or incomplete) firmware.

Quoted 11 lines
root@ubuntu:/sys/devices/pci0000:7c/0000:7c:00.0/0000:7d:00.3# cat irq
0
root@ubuntu:/sys/devices/pci0000:7c/0000:7c:00.0/0000:7d:00.3# ls -l msi_irqs/*
(quoted diff) +0 -3
-r--r--r-- 1 root root 4096 Aug 21 22:04 msi_irqs/499
-r--r--r-- 1 root root 4096 Aug 21 22:04 msi_irqs/500
-r--r--r-- 1 root root 4096 Aug 21 22:04 msi_irqs/501
...
root@ubuntu:/sys/devices/pci0000:7c/0000:7c:00.0/0000:7d:00.3# cat msi_irqs/499
msix
Quoted 1 line

Not quite sure how it is going on different hardware platforms.

My D05 does that as well, and it doesn't expose any INTx support.

Quoted 12 lines
Quoted 11 lines

MSI-X is not something you can "accidentally" use. You have to actively embrace it. In all honesty, this patch tries to move in the wrong direction. If anything, we should kill this hack altogether and fix the (handful of?) drivers that rely on it. That'd actually be a good way to find whether they are still worth keeping in the tree. And if it breaks too many of them, then at least we'll know where we stand.

I'd be tempted to leave the below patch simmer in -next for a few weeks and see if how many people shout:

Quoted 7 lines

This looks like a more proper direction to go.
but here i am wondering how sysfs ABI document should follow the below change doc is patch 2/2:
https://lore.kernel.org/lkml/[email protected]/

On the other hand, my feeling is that nobody should depend on sysfs irq entry nowadays.

Too late. It is there, and we need to preserve it. I just don't think feeding it more erroneous information is the right thing to do.

Quoted 6 lines
Quoted 2 lines

My patch was only dealing with the kernel side of things, not the userspace ABI. That ABI should be carried on unchanged.

it seems this isn't true. your patch is also changing userspace ABI as long as you change pci_dev->irq which will be shown in sysfs irq entry.

I guess I wasn't clear enough above. Let me rephrase this:

My patch was only dealing with the kernel side of things, not the userspace ABI. That ABI should be carried on unchanged, which requires additional changes in the sysfs code.

Quoted 4 lines

if we don't want to change the behaviour of any existing ABI, it seems the only thing we can do here to document it well in ABI doc. i actually doubt anyone has really understood what the irq entry is really showing.

Given that we can't prove that it is actually the case, I believe this is the only option.

Thanks,

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
author
#11 lore raw

Re: [PATCH v2 1/2] PCI/MSI: Fix the confusing IRQ sysfs ABI for MSI-X

Quoted 104 lines · Marc Zyngier
Quoted 94 lines · Barry Song
Quoted 92 lines · Marc Zyngier
Quoted 8 lines · Barry Song
Quoted 6 lines · Marc Zyngier

Hi Bjorn,

Quoted 1 line · Bjorn Helgaas

[...]

Quoted 19 lines
Quoted 19 lines
Quoted 2 lines

In msix_setup_entries(), we get nvecs msi_entry structs, and we get a saved .default_irq in each one?

That's a key point.

Old-school PCI/MSI is represented by a single interrupt, and you *could* somehow make it relatively easy for drivers that only understand INTx to migrate to MSI if you replaced whatever is held in dev->irq (which should only represent the INTx mapping) with the MSI interrupt number. Which I guess is what the MSI code is doing.

This is the 21st century, and nobody should ever rely on such horror, but I'm sure we do have such drivers in the tree. Boo.

However, this *cannot* hold true for Multi-MSI, nor MSI-X, because there is a plurality of interrupts. Even worse, for MSI-X, there is zero guarantee that the allocated interrupts will be in a contiguous space.

Quoted 8 lines
Quoted 3 lines

Given that, what is dev->irq good for? "Absolutely Nothing! (say it again!)".

The only thing is that dev->irq is an sysfs ABI to userspace. Due to the inconsistency between legacy PCI INTx, MSI, MSI-X, this ABI should have been absolutely broken nowadays. This is actually what the patchset was originally aiming at to fix.

I do not think we should expose more of a broken abstraction to userspace. We will have to carry on exposing the first MSI in this field forever, but it doesn't mean we should have to do it for MSI-X.

Quoted 3 lines

One more question from me is that does dev->irq actually hold any valid hardware INTx information while hardware is using MSI-X? At least in my hardware, sysfs ABI for PCI is all "0".

That's probably because nothing actually configured the interrupt, or that there is no INTx implementation. I have that on systems with pretty dodgy (or incomplete) firmware.

Quoted 11 lines
root@ubuntu:/sys/devices/pci0000:7c/0000:7c:00.0/0000:7d:00.3# cat irq
0
root@ubuntu:/sys/devices/pci0000:7c/0000:7c:00.0/0000:7d:00.3# ls -l msi_irqs/*
(quoted diff) +0 -3
-r--r--r-- 1 root root 4096 Aug 21 22:04 msi_irqs/499
-r--r--r-- 1 root root 4096 Aug 21 22:04 msi_irqs/500
-r--r--r-- 1 root root 4096 Aug 21 22:04 msi_irqs/501
...
root@ubuntu:/sys/devices/pci0000:7c/0000:7c:00.0/0000:7d:00.3# cat msi_irqs/499
msix
Quoted 1 line

Not quite sure how it is going on different hardware platforms.

My D05 does that as well, and it doesn't expose any INTx support.

Quoted 12 lines
Quoted 11 lines

MSI-X is not something you can "accidentally" use. You have to actively embrace it. In all honesty, this patch tries to move in the wrong direction. If anything, we should kill this hack altogether and fix the (handful of?) drivers that rely on it. That'd actually be a good way to find whether they are still worth keeping in the tree. And if it breaks too many of them, then at least we'll know where we stand.

I'd be tempted to leave the below patch simmer in -next for a few weeks and see if how many people shout:

Quoted 7 lines

This looks like a more proper direction to go.
but here i am wondering how sysfs ABI document should follow the below change doc is patch 2/2:
https://lore.kernel.org/lkml/[email protected]/

On the other hand, my feeling is that nobody should depend on sysfs irq entry nowadays.

Too late. It is there, and we need to preserve it. I just don't think feeding it more erroneous information is the right thing to do.

Quoted 6 lines
Quoted 2 lines

My patch was only dealing with the kernel side of things, not the userspace ABI. That ABI should be carried on unchanged.

it seems this isn't true. your patch is also changing userspace ABI as long as you change pci_dev->irq which will be shown in sysfs irq entry.

Quoted 5 lines

I guess I wasn't clear enough above. Let me rephrase this:

My patch was only dealing with the kernel side of things, not the userspace ABI. That ABI should be carried on unchanged, which requires additional changes in the sysfs code.

Thanks for clarification. It seems you mean something like below, am I right?

drivers/pci/pci-sysfs.c +18 -1
index 5d63df7..0fa7a16 100644
@@ -26,6 +26,7 @@
 #include <linux/slab.h>
 #include <linux/vgaarb.h>
 #include <linux/pm_runtime.h>
+#include <linux/msi.h>
 #include <linux/of.h>
 #include "pci.h"
 
@@ -49,7 +50,23 @@ static DEVICE_ATTR_RO(field)
 pci_config_attr(subsystem_device, "0x%04x\n");
 pci_config_attr(revision, "0x%02x\n");
 pci_config_attr(class, "0x%06x\n");
-pci_config_attr(irq, "%u\n");
+
+static ssize_t irq_show(struct device *dev,
+                                        struct device_attribute *attr,
+                                        char *buf)
+{
+       struct pci_dev *pdev = to_pci_dev(dev);
+#ifdef CONFIG_PCI_MSI
+       struct msi_desc *desc = first_pci_msi_entry(pdev);
+
+       /* for MSI, return the 1st IRQ in IRQ vector */
+       if (desc && !desc->msi_attrib.is_msix)
+               return sysfs_emit(buf, "%u\n", desc->irq);
+#endif
+
+       return sysfs_emit(buf, "%u\n", pdev->irq);
+}
+static DEVICE_ATTR_RO(irq);
 
 static ssize_t broken_parity_status_show(struct device *dev,
                                         struct device_attribute *attr,
Quoted 8 lines
Quoted 5 lines

if we don't want to change the behaviour of any existing ABI, it seems the only thing we can do here to document it well in ABI doc. i actually doubt anyone has really understood what the irq entry is really showing.

Given that we can't prove that it is actually the case, I believe this is the only option.

we have to document the ABI like below though it seems quite annoying.

1. for devices which don't support MSI and MSI-X, show legacy INTx
2. for devices which support MSI
a. if CONFIG_PCI_MSI is not enabled, show legacy INTx b. if CONFIG_PCI_MSI is enabled and devices are using MSI at this
moment, show 1st IRQ in the vector
c. if CONFIG_PCI_MSI is enabled, but we shutdown its MSI before
the users call sysfs entry,
so at this moment, devices are not using MSI, show legacy INTx
3. for devices which support MSI-X, no matter if it is using MSI-X, show legacy INTx
4. In Addition, INTx might be broken due to incomplete firmware or hardware design for MSI and MSI-X cases

To be honest, it sounds like a disaster :-) but if this is what we have to do, I'd like to try it in v3.

Quoted 7 lines

Thanks,

        M.

--
Without deviation from the norm, progress is not possible.

Thanks
Barry

#12 lore raw

Re: [PATCH v2 1/2] PCI/MSI: Fix the confusing IRQ sysfs ABI for MSI-X

Quoted 11 lines · Barry Song
Quoted 3 lines · Marc Zyngier

On Mon, 23 Aug 2021 12:03:08 +0100,
Barry Song <[email protected]> wrote:

drivers/pci/pci-sysfs.c +6 -0
+static ssize_t irq_show(struct device *dev,
+                                        struct device_attribute *attr,
+                                        char *buf)
+{
+       struct pci_dev *pdev = to_pci_dev(dev);
+#ifdef CONFIG_PCI_MSI

drivers/pci/pci-sysfs.c static DEVICE_ATTR_RO(field) · line 69

Quoted 10 lines
drivers/pci/pci-sysfs.c +10 -0
+       struct msi_desc *desc = first_pci_msi_entry(pdev);
+
+       /* for MSI, return the 1st IRQ in IRQ vector */
+       if (desc && !desc->msi_attrib.is_msix)
+               return sysfs_emit(buf, "%u\n", desc->irq);
+#endif
+
+       return sysfs_emit(buf, "%u\n", pdev->irq);
+}
+static DEVICE_ATTR_RO(irq);

Makes sense to me. And with Marc's patch maybe we could get rid of default_irq, which also seems nice.

Quoted 23 lines
Quoted 7 lines
Quoted 4 lines

if we don't want to change the behaviour of any existing ABI, it seems the only thing we can do here to document it well in ABI doc. i actually doubt anyone has really understood what the irq entry is really showing.

Given that we can't prove that it is actually the case, I believe this is the only option.

we have to document the ABI like below though it seems quite annoying.

1. for devices which don't support MSI and MSI-X, show legacy INTx
2. for devices which support MSI
a. if CONFIG_PCI_MSI is not enabled, show legacy INTx b. if CONFIG_PCI_MSI is enabled and devices are using MSI at this
moment, show 1st IRQ in the vector
c. if CONFIG_PCI_MSI is enabled, but we shutdown its MSI before
the users call sysfs entry,
so at this moment, devices are not using MSI, show legacy INTx
3. for devices which support MSI-X, no matter if it is using MSI-X, show legacy INTx
4. In Addition, INTx might be broken due to incomplete firmware or hardware design for MSI and MSI-X cases

Quoted 2 lines

To be honest, it sounds like a disaster :-) but if this is what we have to do, I'd like to try it in v3.

It doesn't seem necessary to me to get into the gory details of CONFIG_PCI_MSI -- if that's not enabled, drivers can't use MSI anyway.

I don't understand 3. If a device supports both MSI and MSI-X and a driver enables MSI, msi_capability_init() writes dev->irq, so it looks like "irq" should contain the first MSI vector.

I don't understand 4, either. Is the possibility of broken hardware or firmware something we need to document?

What about something like this?

If a driver has enabled MSI (not MSI-X), "irq" contains the IRQ of the first MSI vector. Otherwise "irq" contains the IRQ of the legacy INTx interrupt.

#13 lore raw

Re: [PATCH v2 1/2] PCI/MSI: Fix the confusing IRQ sysfs ABI for MSI-X

Quoted 60 lines · Bjorn Helgaas
Quoted 11 lines · Barry Song
Quoted 3 lines · Marc Zyngier

On Mon, 23 Aug 2021 12:03:08 +0100,
Barry Song <[email protected]> wrote:

drivers/pci/pci-sysfs.c +6 -0
+static ssize_t irq_show(struct device *dev,
+                                        struct device_attribute *attr,
+                                        char *buf)
+{
+       struct pci_dev *pdev = to_pci_dev(dev);
+#ifdef CONFIG_PCI_MSI

drivers/pci/pci-sysfs.c static DEVICE_ATTR_RO(field) · line 69

Quoted 10 lines
drivers/pci/pci-sysfs.c +10 -0
+       struct msi_desc *desc = first_pci_msi_entry(pdev);
+
+       /* for MSI, return the 1st IRQ in IRQ vector */
+       if (desc && !desc->msi_attrib.is_msix)
+               return sysfs_emit(buf, "%u\n", desc->irq);
+#endif
+
+       return sysfs_emit(buf, "%u\n", pdev->irq);
+}
+static DEVICE_ATTR_RO(irq);

Makes sense to me. And with Marc's patch maybe we could get rid of default_irq, which also seems nice.

Quoted 23 lines
Quoted 7 lines
Quoted 4 lines

if we don't want to change the behaviour of any existing ABI, it seems the only thing we can do here to document it well in ABI doc. i actually doubt anyone has really understood what the irq entry is really showing.

Given that we can't prove that it is actually the case, I believe this is the only option.

we have to document the ABI like below though it seems quite annoying.

1. for devices which don't support MSI and MSI-X, show legacy INTx
2. for devices which support MSI
a. if CONFIG_PCI_MSI is not enabled, show legacy INTx b. if CONFIG_PCI_MSI is enabled and devices are using MSI at this
moment, show 1st IRQ in the vector
c. if CONFIG_PCI_MSI is enabled, but we shutdown its MSI before
the users call sysfs entry,
so at this moment, devices are not using MSI, show legacy INTx
3. for devices which support MSI-X, no matter if it is using MSI-X, show legacy INTx
4. In Addition, INTx might be broken due to incomplete firmware or hardware design for MSI and MSI-X cases

Quoted 2 lines

To be honest, it sounds like a disaster :-) but if this is what we have to do, I'd like to try it in v3.

It doesn't seem necessary to me to get into the gory details of CONFIG_PCI_MSI -- if that's not enabled, drivers can't use MSI anyway.

I don't understand 3. If a device supports both MSI and MSI-X and a driver enables MSI, msi_capability_init() writes dev->irq, so it looks like "irq" should contain the first MSI vector.

Quoted 9 lines

I don't understand 4, either. Is the possibility of broken hardware or firmware something we need to document?

What about something like this?

If a driver has enabled MSI (not MSI-X), "irq" contains the IRQ of the first MSI vector. Otherwise "irq" contains the IRQ of the legacy INTx interrupt.

I think that pretty much nails it. CONFIG_MSI is not something that userspace can (nor should) discover anyway.

For (4), you may want to add that

"irq" being set to 0 indicates that the device isn't capable of generating legacy INTx interrupts.

Thanks,

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
author
#14 lore raw

Re: [PATCH v2 1/2] PCI/MSI: Fix the confusing IRQ sysfs ABI for MSI-X

In reply to Marc Zyngier
Quoted 185 lines · Marc Zyngier

Hi Bjorn,

Quoted 90 lines · Bjorn Helgaas

[+cc Thomas, Marc]

Quoted 4 lines · Barry Song

From: Barry Song <[email protected]>

/sys/bus/pci/devices/.../irq sysfs ABI is very confusing at this moment especially for MSI-X cases.

AFAICT this patch *only* affects MSI-X. So are you saying the sysfs ABI is fine for MSI but confusing for MSI-X?

Quoted 27 lines

While MSI sets IRQ to the first
number in the vector, MSI-X does nothing for this though it saves default_irq in msix_setup_entries(). Weird the saved default_irq for MSI-X is never used in pci_msix_shutdown(), which is quite different with pci_msi_shutdown(). Thus, this patch moves to show the first IRQ number which is from the first msi_entry for MSI-X. Hopefully, this can make IRQ ABI more clear and more consistent.

---
 drivers/pci/msi.c | 6 ++++++
 1 file changed, 6 insertions(+)
drivers/pci/msi.c +1 -0
index 9232255..6bbf81b 100644
@@ -771,6 +771,7 @@ static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries,
    int ret;
    u16 control;
    void __iomem *base;
+   struct msi_desc *desc;
 
    /* Ensure MSI-X is disabled while it is set up */
    pci_msix_clear_and_set_ctrl(dev, PCI_MSIX_FLAGS_ENABLE, 0);

drivers/pci/msi.c static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries, · line 820

Quoted 6 lines
drivers/pci/msi.c +3 -0
@@ -814,6 +815,10 @@ static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries,
    pci_msix_clear_and_set_ctrl(dev, PCI_MSIX_FLAGS_MASKALL, 0);
 
    pcibios_free_irq(dev);
+
+   desc = first_pci_msi_entry(dev);
+   dev->irq = desc->irq;

This change is not primarily about sysfs. This is about changing "dev->irq" when MSI-X is enabled, and it's only incidental that sysfs reflects that.

So we need to know the effect of changing dev->irq. Drivers may use the value of dev->irq, and I'm *guessing* this change shouldn't break them since we already do this for MSI, but I'd like some more expert opinion than mine :)

For MSI we have:

  msi_capability_init
    msi_setup_entry
      entry = alloc_msi_entry(nvec)
      entry->msi_attrib.default_irq = dev->irq;     /* Save IOAPIC IRQ */
    dev->irq = entry->irq;

  pci_msi_shutdown
    /* Restore dev->irq to its default pin-assertion IRQ */
    dev->irq = desc->msi_attrib.default_irq;

and for MSI-X we have:

  msix_capability_init
    msix_setup_entries
      for (i = 0; i < nvec; i++)
        entry = alloc_msi_entry(1)
      entry->msi_attrib.default_irq = dev->irq;

pci_msix_shutdown
for_each_pci_msi_entry(entry, dev)
__pci_msix_desc_mask_irq
+ dev->irq = entry->msi_attrib.default_irq; # added by this patch

Things that seem strange to me:

- The msi_setup_entry() comment "Save IOAPIC IRQ" seems needlessly specific; maybe it should be "INTx IRQ".

- The pci_msi_shutdown() comment "Restore ... pin-assertion IRQ" should match the msi_setup_entry() one, e.g., maybe it should also be "INTx IRQ". There are no INTx or IOAPIC pins in PCIe.

Quoted 8 lines

- The only use of .default_irq is to save and restore dev->irq, so it looks like a per-device thing, not a per-vector thing.

In msi_setup_entry() there's only one msi_entry, so there's only one saved .default_irq.

In msix_setup_entries(), we get nvecs msi_entry structs, and we get a saved .default_irq in each one?

That's a key point.

Old-school PCI/MSI is represented by a single interrupt, and you *could* somehow make it relatively easy for drivers that only understand INTx to migrate to MSI if you replaced whatever is held in dev->irq (which should only represent the INTx mapping) with the MSI interrupt number. Which I guess is what the MSI code is doing.

This is the 21st century, and nobody should ever rely on such horror, but I'm sure we do have such drivers in the tree. Boo.

However, this *cannot* hold true for Multi-MSI, nor MSI-X, because there is a plurality of interrupts. Even worse, for MSI-X, there is zero guarantee that the allocated interrupts will be in a contiguous space.

Given that, what is dev->irq good for? "Absolutely Nothing! (say it again!)".

MSI-X is not something you can "accidentally" use. You have to actively embrace it. In all honesty, this patch tries to move in the wrong direction. If anything, we should kill this hack altogether and fix the (handful of?) drivers that rely on it. That'd actually be a good way to find whether they are still worth keeping in the tree. And if it breaks too many of them, then at least we'll know where we stand.

I'd be tempted to leave the below patch simmer in -next for a few weeks and see if how many people shout:

drivers/pci/msi.c +0 -5
index e5e75331b415..2be9a01cbe72 100644
@@ -591,7 +591,6 @@ msi_setup_entry(struct pci_dev *dev, int nvec, struct irq_affinity *affd)
        entry->msi_attrib.is_virtual    = 0;
        entry->msi_attrib.entry_nr      = 0;
        entry->msi_attrib.maskbit       = !!(control & PCI_MSI_FLAGS_MASKBIT);
-       entry->msi_attrib.default_irq   = dev->irq;     /* Save IOAPIC IRQ */
        entry->msi_attrib.multi_cap     = (control & PCI_MSI_FLAGS_QMASK) >> 1;
        entry->msi_attrib.multiple      = ilog2(__roundup_pow_of_two(nvec));
 
@@ -682,7 +681,6 @@ static int msi_capability_init(struct pci_dev *dev, int nvec,
        dev->msi_enabled = 1;
 
        pcibios_free_irq(dev);
-       dev->irq = entry->irq;
        return 0;
 }
 
@@ -742,7 +740,6 @@ static int msix_setup_entries(struct pci_dev *dev, void __iomem *base,
                entry->msi_attrib.is_virtual =
                        entry->msi_attrib.entry_nr >= vec_count;
 
-               entry->msi_attrib.default_irq   = dev->irq;
                entry->mask_base                = base;
 
                addr = pci_msix_desc_addr(entry);
@@ -964,8 +961,6 @@ static void pci_msi_shutdown(struct pci_dev *dev)
        mask = msi_mask(desc->msi_attrib.multi_cap);
        msi_mask_irq(desc, mask, 0);
 
-       /* Restore dev->irq to its default pin-assertion IRQ */
-       dev->irq = desc->msi_attrib.default_irq;
        pcibios_alloc_irq(dev);
 }
 
include/linux/msi.h +0 -1
index e8bdcb83172b..a631664c1c38 100644
@@ -114,7 +114,6 @@ struct ti_sci_inta_msi_desc {
  * @maskbit:   [PCI MSI/X] Mask-Pending bit supported?
  * @is_64:     [PCI MSI/X] Address size: 0=32bit 1=64bit
  * @entry_nr:  [PCI MSI/X] Entry which is described by this descriptor
- * @default_irq:[PCI MSI/X] The default pre-assigned non-MSI irq
  * @mask_pos:  [PCI MSI]   Mask register position
  * @mask_base: [PCI MSI-X] Mask register base address
  * @platform:  [platform]  Platform device specific msi descriptor data

include/linux/msi.h struct msi_desc { · line 152

Quoted 7 lines
include/linux/msi.h +0 -1
@@ -148,7 +147,6 @@ struct msi_desc {
                                u8      is_64           : 1;
                                u8      is_virtual      : 1;
                                u16     entry_nr;
-                               unsigned default_irq;
                        } msi_attrib;
                        union {
                                u8      mask_pos;

We will also need the below change as pci_irq_vector() depends on dev->irq for the MSI case.

int pci_irq_vector(struct pci_dev *dev, unsigned int nr)

{
        if (dev->msix_enabled) {
                struct msi_desc *entry;
                int i = 0;

                for_each_pci_msi_entry(entry, dev) {
                        if (i == nr)
                                return entry->irq;
                        i++;
                }
                WARN_ON_ONCE(1);
                return -EINVAL;
        }

        if (dev->msi_enabled) {
                struct msi_desc *entry = first_pci_msi_entry(dev);

                if (WARN_ON_ONCE(nr >= entry->nvec_used))
                        return -EINVAL;
+                return entry->irq + nr;
        } else {
                if (WARN_ON_ONCE(nr > 0))
                        return -EINVAL;
        }

- return dev->irq + nr;
+ return dev->irq;
}
EXPORT_SYMBOL(pci_irq_vector);

Quoted 6 lines

Thanks,

        M.

--
Without deviation from the norm, progress is not possible.

Thanks
barry

author
#15 lore raw

Re: [PATCH v2 1/2] PCI/MSI: Fix the confusing IRQ sysfs ABI for MSI-X

Quoted 228 lines · Barry Song
Quoted 185 lines · Marc Zyngier

Hi Bjorn,

Quoted 90 lines · Bjorn Helgaas

[+cc Thomas, Marc]

Quoted 4 lines · Barry Song

From: Barry Song <[email protected]>

/sys/bus/pci/devices/.../irq sysfs ABI is very confusing at this moment especially for MSI-X cases.

AFAICT this patch *only* affects MSI-X. So are you saying the sysfs ABI is fine for MSI but confusing for MSI-X?

Quoted 27 lines

While MSI sets IRQ to the first
number in the vector, MSI-X does nothing for this though it saves default_irq in msix_setup_entries(). Weird the saved default_irq for MSI-X is never used in pci_msix_shutdown(), which is quite different with pci_msi_shutdown(). Thus, this patch moves to show the first IRQ number which is from the first msi_entry for MSI-X. Hopefully, this can make IRQ ABI more clear and more consistent.

---
 drivers/pci/msi.c | 6 ++++++
 1 file changed, 6 insertions(+)
drivers/pci/msi.c +1 -0
index 9232255..6bbf81b 100644
@@ -771,6 +771,7 @@ static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries,
    int ret;
    u16 control;
    void __iomem *base;
+   struct msi_desc *desc;
 
    /* Ensure MSI-X is disabled while it is set up */
    pci_msix_clear_and_set_ctrl(dev, PCI_MSIX_FLAGS_ENABLE, 0);

drivers/pci/msi.c static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries, · line 820

Quoted 6 lines
drivers/pci/msi.c +3 -0
@@ -814,6 +815,10 @@ static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries,
    pci_msix_clear_and_set_ctrl(dev, PCI_MSIX_FLAGS_MASKALL, 0);
 
    pcibios_free_irq(dev);
+
+   desc = first_pci_msi_entry(dev);
+   dev->irq = desc->irq;

This change is not primarily about sysfs. This is about changing "dev->irq" when MSI-X is enabled, and it's only incidental that sysfs reflects that.

So we need to know the effect of changing dev->irq. Drivers may use the value of dev->irq, and I'm *guessing* this change shouldn't break them since we already do this for MSI, but I'd like some more expert opinion than mine :)

For MSI we have:

  msi_capability_init
    msi_setup_entry
      entry = alloc_msi_entry(nvec)
      entry->msi_attrib.default_irq = dev->irq;     /* Save IOAPIC IRQ */
    dev->irq = entry->irq;

  pci_msi_shutdown
    /* Restore dev->irq to its default pin-assertion IRQ */
    dev->irq = desc->msi_attrib.default_irq;

and for MSI-X we have:

  msix_capability_init
    msix_setup_entries
      for (i = 0; i < nvec; i++)
        entry = alloc_msi_entry(1)
      entry->msi_attrib.default_irq = dev->irq;

pci_msix_shutdown
for_each_pci_msi_entry(entry, dev)
__pci_msix_desc_mask_irq
+ dev->irq = entry->msi_attrib.default_irq; # added by this patch

Things that seem strange to me:

- The msi_setup_entry() comment "Save IOAPIC IRQ" seems needlessly specific; maybe it should be "INTx IRQ".

- The pci_msi_shutdown() comment "Restore ... pin-assertion IRQ" should match the msi_setup_entry() one, e.g., maybe it should also be "INTx IRQ". There are no INTx or IOAPIC pins in PCIe.

Quoted 8 lines

- The only use of .default_irq is to save and restore dev->irq, so it looks like a per-device thing, not a per-vector thing.

In msi_setup_entry() there's only one msi_entry, so there's only one saved .default_irq.

In msix_setup_entries(), we get nvecs msi_entry structs, and we get a saved .default_irq in each one?

That's a key point.

Old-school PCI/MSI is represented by a single interrupt, and you *could* somehow make it relatively easy for drivers that only understand INTx to migrate to MSI if you replaced whatever is held in dev->irq (which should only represent the INTx mapping) with the MSI interrupt number. Which I guess is what the MSI code is doing.

This is the 21st century, and nobody should ever rely on such horror, but I'm sure we do have such drivers in the tree. Boo.

However, this *cannot* hold true for Multi-MSI, nor MSI-X, because there is a plurality of interrupts. Even worse, for MSI-X, there is zero guarantee that the allocated interrupts will be in a contiguous space.

Given that, what is dev->irq good for? "Absolutely Nothing! (say it again!)".

MSI-X is not something you can "accidentally" use. You have to actively embrace it. In all honesty, this patch tries to move in the wrong direction. If anything, we should kill this hack altogether and fix the (handful of?) drivers that rely on it. That'd actually be a good way to find whether they are still worth keeping in the tree. And if it breaks too many of them, then at least we'll know where we stand.

I'd be tempted to leave the below patch simmer in -next for a few weeks and see if how many people shout:

drivers/pci/msi.c +0 -5
index e5e75331b415..2be9a01cbe72 100644
@@ -591,7 +591,6 @@ msi_setup_entry(struct pci_dev *dev, int nvec, struct irq_affinity *affd)
        entry->msi_attrib.is_virtual    = 0;
        entry->msi_attrib.entry_nr      = 0;
        entry->msi_attrib.maskbit       = !!(control & PCI_MSI_FLAGS_MASKBIT);
-       entry->msi_attrib.default_irq   = dev->irq;     /* Save IOAPIC IRQ */
        entry->msi_attrib.multi_cap     = (control & PCI_MSI_FLAGS_QMASK) >> 1;
        entry->msi_attrib.multiple      = ilog2(__roundup_pow_of_two(nvec));
 
@@ -682,7 +681,6 @@ static int msi_capability_init(struct pci_dev *dev, int nvec,
        dev->msi_enabled = 1;
 
        pcibios_free_irq(dev);
-       dev->irq = entry->irq;
        return 0;
 }
 
@@ -742,7 +740,6 @@ static int msix_setup_entries(struct pci_dev *dev, void __iomem *base,
                entry->msi_attrib.is_virtual =
                        entry->msi_attrib.entry_nr >= vec_count;
 
-               entry->msi_attrib.default_irq   = dev->irq;
                entry->mask_base                = base;
 
                addr = pci_msix_desc_addr(entry);
@@ -964,8 +961,6 @@ static void pci_msi_shutdown(struct pci_dev *dev)
        mask = msi_mask(desc->msi_attrib.multi_cap);
        msi_mask_irq(desc, mask, 0);
 
-       /* Restore dev->irq to its default pin-assertion IRQ */
-       dev->irq = desc->msi_attrib.default_irq;
        pcibios_alloc_irq(dev);
 }
 
include/linux/msi.h +0 -1
index e8bdcb83172b..a631664c1c38 100644
@@ -114,7 +114,6 @@ struct ti_sci_inta_msi_desc {
  * @maskbit:   [PCI MSI/X] Mask-Pending bit supported?
  * @is_64:     [PCI MSI/X] Address size: 0=32bit 1=64bit
  * @entry_nr:  [PCI MSI/X] Entry which is described by this descriptor
- * @default_irq:[PCI MSI/X] The default pre-assigned non-MSI irq
  * @mask_pos:  [PCI MSI]   Mask register position
  * @mask_base: [PCI MSI-X] Mask register base address
  * @platform:  [platform]  Platform device specific msi descriptor data

include/linux/msi.h struct msi_desc { · line 152

Quoted 7 lines
include/linux/msi.h +0 -1
@@ -148,7 +147,6 @@ struct msi_desc {
                                u8      is_64           : 1;
                                u8      is_virtual      : 1;
                                u16     entry_nr;
-                               unsigned default_irq;
                        } msi_attrib;
                        union {
                                u8      mask_pos;

We will also need the below change as pci_irq_vector() depends on dev->irq for the MSI case.

int pci_irq_vector(struct pci_dev *dev, unsigned int nr)

{
        if (dev->msix_enabled) {
                struct msi_desc *entry;
                int i = 0;

                for_each_pci_msi_entry(entry, dev) {
                        if (i == nr)
                                return entry->irq;
                        i++;
                }
                WARN_ON_ONCE(1);
                return -EINVAL;
        }

        if (dev->msi_enabled) {
                struct msi_desc *entry = first_pci_msi_entry(dev);

                if (WARN_ON_ONCE(nr >= entry->nvec_used))
                        return -EINVAL;
(quoted diff) +2 -1
+                return entry->irq + nr;
        } else {
                if (WARN_ON_ONCE(nr > 0))
                        return -EINVAL;
        }
 
 
-        return dev->irq + nr;
+       return dev->irq;
Quoted 3 lines

}
EXPORT_SYMBOL(pci_irq_vector);

And here:

drivers/pci/msi.c +1 -1
@@ -401,7 +401,7 @@ static void __pci_restore_msi_state(struct pci_dev *dev)
        if (!dev->msi_enabled)
                return;
 
-       entry = irq_get_msi_desc(dev->irq);
+       entry = first_pci_msi_entry(dev);
 
        pci_intx_for_msi(dev, 0);
        pci_msi_set_enable(dev, 0);

since drivers/net/ethernet/intel/ice/ice_virtchnl_pf.c is the only one calling pci_restore_msi_state(),
will probably require one test from the driver maintainers.

Quoted 7 lines
Quoted 6 lines

Thanks,

        M.

--
Without deviation from the norm, progress is not possible.

Thanks
barry

#16 lore raw

Re: [PATCH v2 1/2] PCI/MSI: Fix the confusing IRQ sysfs ABI for MSI-X

Quoted 248 lines · Barry Song
Quoted 228 lines · Barry Song
Quoted 185 lines · Marc Zyngier

Hi Bjorn,

Quoted 90 lines · Bjorn Helgaas

[+cc Thomas, Marc]

Quoted 4 lines · Barry Song

From: Barry Song <[email protected]>

/sys/bus/pci/devices/.../irq sysfs ABI is very confusing at this moment especially for MSI-X cases.

AFAICT this patch *only* affects MSI-X. So are you saying the sysfs ABI is fine for MSI but confusing for MSI-X?

Quoted 27 lines

While MSI sets IRQ to the first
number in the vector, MSI-X does nothing for this though it saves default_irq in msix_setup_entries(). Weird the saved default_irq for MSI-X is never used in pci_msix_shutdown(), which is quite different with pci_msi_shutdown(). Thus, this patch moves to show the first IRQ number which is from the first msi_entry for MSI-X. Hopefully, this can make IRQ ABI more clear and more consistent.

---
 drivers/pci/msi.c | 6 ++++++
 1 file changed, 6 insertions(+)
drivers/pci/msi.c +1 -0
index 9232255..6bbf81b 100644
@@ -771,6 +771,7 @@ static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries,
    int ret;
    u16 control;
    void __iomem *base;
+   struct msi_desc *desc;
 
    /* Ensure MSI-X is disabled while it is set up */
    pci_msix_clear_and_set_ctrl(dev, PCI_MSIX_FLAGS_ENABLE, 0);

drivers/pci/msi.c static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries, · line 820

Quoted 6 lines
drivers/pci/msi.c +3 -0
@@ -814,6 +815,10 @@ static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries,
    pci_msix_clear_and_set_ctrl(dev, PCI_MSIX_FLAGS_MASKALL, 0);
 
    pcibios_free_irq(dev);
+
+   desc = first_pci_msi_entry(dev);
+   dev->irq = desc->irq;

This change is not primarily about sysfs. This is about changing "dev->irq" when MSI-X is enabled, and it's only incidental that sysfs reflects that.

So we need to know the effect of changing dev->irq. Drivers may use the value of dev->irq, and I'm *guessing* this change shouldn't break them since we already do this for MSI, but I'd like some more expert opinion than mine :)

For MSI we have:

  msi_capability_init
    msi_setup_entry
      entry = alloc_msi_entry(nvec)
      entry->msi_attrib.default_irq = dev->irq;     /* Save IOAPIC IRQ */
    dev->irq = entry->irq;

  pci_msi_shutdown
    /* Restore dev->irq to its default pin-assertion IRQ */
    dev->irq = desc->msi_attrib.default_irq;

and for MSI-X we have:

  msix_capability_init
    msix_setup_entries
      for (i = 0; i < nvec; i++)
        entry = alloc_msi_entry(1)
      entry->msi_attrib.default_irq = dev->irq;

pci_msix_shutdown
for_each_pci_msi_entry(entry, dev)
__pci_msix_desc_mask_irq
+ dev->irq = entry->msi_attrib.default_irq; # added by this patch

Things that seem strange to me:

- The msi_setup_entry() comment "Save IOAPIC IRQ" seems needlessly specific; maybe it should be "INTx IRQ".

- The pci_msi_shutdown() comment "Restore ... pin-assertion IRQ" should match the msi_setup_entry() one, e.g., maybe it should also be "INTx IRQ". There are no INTx or IOAPIC pins in PCIe.

Quoted 8 lines

- The only use of .default_irq is to save and restore dev->irq, so it looks like a per-device thing, not a per-vector thing.

In msi_setup_entry() there's only one msi_entry, so there's only one saved .default_irq.

In msix_setup_entries(), we get nvecs msi_entry structs, and we get a saved .default_irq in each one?

That's a key point.

Old-school PCI/MSI is represented by a single interrupt, and you *could* somehow make it relatively easy for drivers that only understand INTx to migrate to MSI if you replaced whatever is held in dev->irq (which should only represent the INTx mapping) with the MSI interrupt number. Which I guess is what the MSI code is doing.

This is the 21st century, and nobody should ever rely on such horror, but I'm sure we do have such drivers in the tree. Boo.

However, this *cannot* hold true for Multi-MSI, nor MSI-X, because there is a plurality of interrupts. Even worse, for MSI-X, there is zero guarantee that the allocated interrupts will be in a contiguous space.

Given that, what is dev->irq good for? "Absolutely Nothing! (say it again!)".

MSI-X is not something you can "accidentally" use. You have to actively embrace it. In all honesty, this patch tries to move in the wrong direction. If anything, we should kill this hack altogether and fix the (handful of?) drivers that rely on it. That'd actually be a good way to find whether they are still worth keeping in the tree. And if it breaks too many of them, then at least we'll know where we stand.

I'd be tempted to leave the below patch simmer in -next for a few weeks and see if how many people shout:

drivers/pci/msi.c +0 -5
index e5e75331b415..2be9a01cbe72 100644
@@ -591,7 +591,6 @@ msi_setup_entry(struct pci_dev *dev, int nvec, struct irq_affinity *affd)
        entry->msi_attrib.is_virtual    = 0;
        entry->msi_attrib.entry_nr      = 0;
        entry->msi_attrib.maskbit       = !!(control & PCI_MSI_FLAGS_MASKBIT);
-       entry->msi_attrib.default_irq   = dev->irq;     /* Save IOAPIC IRQ */
        entry->msi_attrib.multi_cap     = (control & PCI_MSI_FLAGS_QMASK) >> 1;
        entry->msi_attrib.multiple      = ilog2(__roundup_pow_of_two(nvec));
 
@@ -682,7 +681,6 @@ static int msi_capability_init(struct pci_dev *dev, int nvec,
        dev->msi_enabled = 1;
 
        pcibios_free_irq(dev);
-       dev->irq = entry->irq;
        return 0;
 }
 
@@ -742,7 +740,6 @@ static int msix_setup_entries(struct pci_dev *dev, void __iomem *base,
                entry->msi_attrib.is_virtual =
                        entry->msi_attrib.entry_nr >= vec_count;
 
-               entry->msi_attrib.default_irq   = dev->irq;
                entry->mask_base                = base;
 
                addr = pci_msix_desc_addr(entry);
@@ -964,8 +961,6 @@ static void pci_msi_shutdown(struct pci_dev *dev)
        mask = msi_mask(desc->msi_attrib.multi_cap);
        msi_mask_irq(desc, mask, 0);
 
-       /* Restore dev->irq to its default pin-assertion IRQ */
-       dev->irq = desc->msi_attrib.default_irq;
        pcibios_alloc_irq(dev);
 }
 
include/linux/msi.h +0 -1
index e8bdcb83172b..a631664c1c38 100644
@@ -114,7 +114,6 @@ struct ti_sci_inta_msi_desc {
  * @maskbit:   [PCI MSI/X] Mask-Pending bit supported?
  * @is_64:     [PCI MSI/X] Address size: 0=32bit 1=64bit
  * @entry_nr:  [PCI MSI/X] Entry which is described by this descriptor
- * @default_irq:[PCI MSI/X] The default pre-assigned non-MSI irq
  * @mask_pos:  [PCI MSI]   Mask register position
  * @mask_base: [PCI MSI-X] Mask register base address
  * @platform:  [platform]  Platform device specific msi descriptor data

include/linux/msi.h struct msi_desc { · line 152

Quoted 7 lines
include/linux/msi.h +0 -1
@@ -148,7 +147,6 @@ struct msi_desc {
                                u8      is_64           : 1;
                                u8      is_virtual      : 1;
                                u16     entry_nr;
-                               unsigned default_irq;
                        } msi_attrib;
                        union {
                                u8      mask_pos;

We will also need the below change as pci_irq_vector() depends on dev->irq for the MSI case.

int pci_irq_vector(struct pci_dev *dev, unsigned int nr)

{
        if (dev->msix_enabled) {
                struct msi_desc *entry;
                int i = 0;

                for_each_pci_msi_entry(entry, dev) {
                        if (i == nr)
                                return entry->irq;
                        i++;
                }
                WARN_ON_ONCE(1);
                return -EINVAL;
        }

        if (dev->msi_enabled) {
                struct msi_desc *entry = first_pci_msi_entry(dev);

                if (WARN_ON_ONCE(nr >= entry->nvec_used))
                        return -EINVAL;
(quoted diff) +2 -1
+                return entry->irq + nr;
        } else {
                if (WARN_ON_ONCE(nr > 0))
                        return -EINVAL;
        }
 
 
-        return dev->irq + nr;
+       return dev->irq;
Quoted 3 lines

}
EXPORT_SYMBOL(pci_irq_vector);

And here:

drivers/pci/msi.c +1 -1
@@ -401,7 +401,7 @@ static void __pci_restore_msi_state(struct pci_dev *dev)
        if (!dev->msi_enabled)
                return;
 
-       entry = irq_get_msi_desc(dev->irq);
+       entry = first_pci_msi_entry(dev);
 
        pci_intx_for_msi(dev, 0);
        pci_msi_set_enable(dev, 0);
Quoted 3 lines

since drivers/net/ethernet/intel/ice/ice_virtchnl_pf.c is the only one calling pci_restore_msi_state(), will probably require one test from the driver maintainers.

All good catches. This needs tons of testing, and I fully expect regressions (at least drivers falling back to INTx when they were planning to use MSIs).

I'll try to roll a proper patch once 5.14 is out.

Thanks,

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
author
#17 lore raw

Re: [PATCH v2 1/2] PCI/MSI: Fix the confusing IRQ sysfs ABI for MSI-X

Quoted 259 lines · Marc Zyngier
Quoted 248 lines · Barry Song
Quoted 228 lines · Barry Song
Quoted 185 lines · Marc Zyngier

Hi Bjorn,

Quoted 90 lines · Bjorn Helgaas

[+cc Thomas, Marc]

Quoted 4 lines · Barry Song

From: Barry Song <[email protected]>

/sys/bus/pci/devices/.../irq sysfs ABI is very confusing at this moment especially for MSI-X cases.

AFAICT this patch *only* affects MSI-X. So are you saying the sysfs ABI is fine for MSI but confusing for MSI-X?

Quoted 27 lines

While MSI sets IRQ to the first
number in the vector, MSI-X does nothing for this though it saves default_irq in msix_setup_entries(). Weird the saved default_irq for MSI-X is never used in pci_msix_shutdown(), which is quite different with pci_msi_shutdown(). Thus, this patch moves to show the first IRQ number which is from the first msi_entry for MSI-X. Hopefully, this can make IRQ ABI more clear and more consistent.

---
 drivers/pci/msi.c | 6 ++++++
 1 file changed, 6 insertions(+)
drivers/pci/msi.c +1 -0
index 9232255..6bbf81b 100644
@@ -771,6 +771,7 @@ static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries,
    int ret;
    u16 control;
    void __iomem *base;
+   struct msi_desc *desc;
 
    /* Ensure MSI-X is disabled while it is set up */
    pci_msix_clear_and_set_ctrl(dev, PCI_MSIX_FLAGS_ENABLE, 0);

drivers/pci/msi.c static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries, · line 820

Quoted 6 lines
drivers/pci/msi.c +3 -0
@@ -814,6 +815,10 @@ static int msix_capability_init(struct pci_dev *dev, struct msix_entry *entries,
    pci_msix_clear_and_set_ctrl(dev, PCI_MSIX_FLAGS_MASKALL, 0);
 
    pcibios_free_irq(dev);
+
+   desc = first_pci_msi_entry(dev);
+   dev->irq = desc->irq;

This change is not primarily about sysfs. This is about changing "dev->irq" when MSI-X is enabled, and it's only incidental that sysfs reflects that.

So we need to know the effect of changing dev->irq. Drivers may use the value of dev->irq, and I'm *guessing* this change shouldn't break them since we already do this for MSI, but I'd like some more expert opinion than mine :)

For MSI we have:

  msi_capability_init
    msi_setup_entry
      entry = alloc_msi_entry(nvec)
      entry->msi_attrib.default_irq = dev->irq;     /* Save IOAPIC IRQ */
    dev->irq = entry->irq;

  pci_msi_shutdown
    /* Restore dev->irq to its default pin-assertion IRQ */
    dev->irq = desc->msi_attrib.default_irq;

and for MSI-X we have:

  msix_capability_init
    msix_setup_entries
      for (i = 0; i < nvec; i++)
        entry = alloc_msi_entry(1)
      entry->msi_attrib.default_irq = dev->irq;

pci_msix_shutdown
for_each_pci_msi_entry(entry, dev)
__pci_msix_desc_mask_irq
+ dev->irq = entry->msi_attrib.default_irq; # added by this patch

Things that seem strange to me:

- The msi_setup_entry() comment "Save IOAPIC IRQ" seems needlessly specific; maybe it should be "INTx IRQ".

- The pci_msi_shutdown() comment "Restore ... pin-assertion IRQ" should match the msi_setup_entry() one, e.g., maybe it should also be "INTx IRQ". There are no INTx or IOAPIC pins in PCIe.

Quoted 8 lines

- The only use of .default_irq is to save and restore dev->irq, so it looks like a per-device thing, not a per-vector thing.

In msi_setup_entry() there's only one msi_entry, so there's only one saved .default_irq.

In msix_setup_entries(), we get nvecs msi_entry structs, and we get a saved .default_irq in each one?

That's a key point.

Old-school PCI/MSI is represented by a single interrupt, and you *could* somehow make it relatively easy for drivers that only understand INTx to migrate to MSI if you replaced whatever is held in dev->irq (which should only represent the INTx mapping) with the MSI interrupt number. Which I guess is what the MSI code is doing.

This is the 21st century, and nobody should ever rely on such horror, but I'm sure we do have such drivers in the tree. Boo.

However, this *cannot* hold true for Multi-MSI, nor MSI-X, because there is a plurality of interrupts. Even worse, for MSI-X, there is zero guarantee that the allocated interrupts will be in a contiguous space.

Given that, what is dev->irq good for? "Absolutely Nothing! (say it again!)".

MSI-X is not something you can "accidentally" use. You have to actively embrace it. In all honesty, this patch tries to move in the wrong direction. If anything, we should kill this hack altogether and fix the (handful of?) drivers that rely on it. That'd actually be a good way to find whether they are still worth keeping in the tree. And if it breaks too many of them, then at least we'll know where we stand.

I'd be tempted to leave the below patch simmer in -next for a few weeks and see if how many people shout:

drivers/pci/msi.c +0 -5
index e5e75331b415..2be9a01cbe72 100644
@@ -591,7 +591,6 @@ msi_setup_entry(struct pci_dev *dev, int nvec, struct irq_affinity *affd)
        entry->msi_attrib.is_virtual    = 0;
        entry->msi_attrib.entry_nr      = 0;
        entry->msi_attrib.maskbit       = !!(control & PCI_MSI_FLAGS_MASKBIT);
-       entry->msi_attrib.default_irq   = dev->irq;     /* Save IOAPIC IRQ */
        entry->msi_attrib.multi_cap     = (control & PCI_MSI_FLAGS_QMASK) >> 1;
        entry->msi_attrib.multiple      = ilog2(__roundup_pow_of_two(nvec));
 
@@ -682,7 +681,6 @@ static int msi_capability_init(struct pci_dev *dev, int nvec,
        dev->msi_enabled = 1;
 
        pcibios_free_irq(dev);
-       dev->irq = entry->irq;
        return 0;
 }
 
@@ -742,7 +740,6 @@ static int msix_setup_entries(struct pci_dev *dev, void __iomem *base,
                entry->msi_attrib.is_virtual =
                        entry->msi_attrib.entry_nr >= vec_count;
 
-               entry->msi_attrib.default_irq   = dev->irq;
                entry->mask_base                = base;
 
                addr = pci_msix_desc_addr(entry);
@@ -964,8 +961,6 @@ static void pci_msi_shutdown(struct pci_dev *dev)
        mask = msi_mask(desc->msi_attrib.multi_cap);
        msi_mask_irq(desc, mask, 0);
 
-       /* Restore dev->irq to its default pin-assertion IRQ */
-       dev->irq = desc->msi_attrib.default_irq;
        pcibios_alloc_irq(dev);
 }
 
include/linux/msi.h +0 -1
index e8bdcb83172b..a631664c1c38 100644
@@ -114,7 +114,6 @@ struct ti_sci_inta_msi_desc {
  * @maskbit:   [PCI MSI/X] Mask-Pending bit supported?
  * @is_64:     [PCI MSI/X] Address size: 0=32bit 1=64bit
  * @entry_nr:  [PCI MSI/X] Entry which is described by this descriptor
- * @default_irq:[PCI MSI/X] The default pre-assigned non-MSI irq
  * @mask_pos:  [PCI MSI]   Mask register position
  * @mask_base: [PCI MSI-X] Mask register base address
  * @platform:  [platform]  Platform device specific msi descriptor data

include/linux/msi.h struct msi_desc { · line 152

Quoted 7 lines
include/linux/msi.h +0 -1
@@ -148,7 +147,6 @@ struct msi_desc {
                                u8      is_64           : 1;
                                u8      is_virtual      : 1;
                                u16     entry_nr;
-                               unsigned default_irq;
                        } msi_attrib;
                        union {
                                u8      mask_pos;

We will also need the below change as pci_irq_vector() depends on dev->irq for the MSI case.

int pci_irq_vector(struct pci_dev *dev, unsigned int nr)

{
        if (dev->msix_enabled) {
                struct msi_desc *entry;
                int i = 0;

                for_each_pci_msi_entry(entry, dev) {
                        if (i == nr)
                                return entry->irq;
                        i++;
                }
                WARN_ON_ONCE(1);
                return -EINVAL;
        }

        if (dev->msi_enabled) {
                struct msi_desc *entry = first_pci_msi_entry(dev);

                if (WARN_ON_ONCE(nr >= entry->nvec_used))
                        return -EINVAL;
(quoted diff) +2 -1
+                return entry->irq + nr;
        } else {
                if (WARN_ON_ONCE(nr > 0))
                        return -EINVAL;
        }
 
 
-        return dev->irq + nr;
+       return dev->irq;
Quoted 3 lines

}
EXPORT_SYMBOL(pci_irq_vector);

And here:

drivers/pci/msi.c +1 -1
@@ -401,7 +401,7 @@ static void __pci_restore_msi_state(struct pci_dev *dev)
        if (!dev->msi_enabled)
                return;
 
-       entry = irq_get_msi_desc(dev->irq);
+       entry = first_pci_msi_entry(dev);
 
        pci_intx_for_msi(dev, 0);
        pci_msi_set_enable(dev, 0);
Quoted 3 lines

since drivers/net/ethernet/intel/ice/ice_virtchnl_pf.c is the only one calling pci_restore_msi_state(), will probably require one test from the driver maintainers.

All good catches. This needs tons of testing, and I fully expect regressions (at least drivers falling back to INTx when they were planning to use MSIs).

Quoted 1 line

I'll try to roll a proper patch once 5.14 is out.

Thanks! Anyway, I've integrated all good suggestions from the discussion into v3:
https://lore.kernel.org/lkml/[email protected]/
So that we can have a new base to start after 5.14 is out.

Quoted 7 lines

Thanks,

        M.

--
Without deviation from the norm, progress is not possible.

Thanks
barry

author
#18 lore raw

PATCH v2 2/2 Documentation: ABI: sysfs-bus-pci: Add description for IRQ entry

In reply to Barry Song

From: Barry Song <[email protected]>

/sys/bus/pci/devices/.../irq has been there for many years but it has never been documented. This patch is trying to document it.

---
 Documentation/ABI/testing/sysfs-bus-pci | 8 ++++++++
 1 file changed, 8 insertions(+)
Documentation/ABI/testing/sysfs-bus-pci +8 -0
index 793cbb7..8d42385 100644
@@ -96,6 +96,14 @@ Description:
 		This attribute indicates the mode that the irq vector named by
 		the file is in (msi vs. msix)
 
+What:		/sys/bus/pci/devices/.../irq
+Date:		August 2021
+Contact:	Barry Song <[email protected]>
+Description:
+		Historically this attribute represent the IRQ line which runs
+		from the PCI device to the Interrupt controller. With MSI and
+		MSI-X, this attribute is the first IRQ number of IRQ vectors.
+
 What:		/sys/bus/pci/devices/.../remove
 Date:		January 2009
 Contact:	Linux PCI developers <[email protected]>
Signature · 1 line
1.8.3.1