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

On Wed, Aug 25, 2021 at 10:24 PM Marc Zyngier [off-list ref] wrote:

On Tue, 24 Aug 2021 22:29:59 +0100,
Barry Song [off-list ref] wrote:

quoted

On Wed, Aug 25, 2021 at 8:51 AM Barry Song [off-list ref] wrote:

quoted

On Sat, Aug 21, 2021 at 10:42 PM Marc Zyngier [off-list ref] wrote:

quoted

Hi Bjorn,

On Sat, 21 Aug 2021 00:33:28 +0100,
Bjorn Helgaas [off-list ref] wrote:

quoted

[+cc Thomas, Marc]

On Sat, Aug 21, 2021 at 10:37:43AM +1200, Barry Song wrote:

quoted

From: Barry Song <redacted>

/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

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.

Acked-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Signed-off-by: Barry Song <redacted>
---
 drivers/pci/msi.c | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/drivers/pci/msi.c b/drivers/pci/msi.c
index 9232255..6bbf81b 100644
--- a/drivers/pci/msi.c
+++ b/drivers/pci/msi.c

@@ -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;

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?

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:

diff --git a/drivers/pci/msi.c b/drivers/pci/msi.c
index e5e75331b415..2be9a01cbe72 100644
--- a/drivers/pci/msi.c
+++ b/drivers/pci/msi.c

@@ -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);
 }

diff --git a/include/linux/msi.h b/include/linux/msi.h
index e8bdcb83172b..a631664c1c38 100644
--- a/include/linux/msi.h
+++ b/include/linux/msi.h

@@ -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;

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);

And here:

--- a/drivers/pci/msi.c
+++ b/drivers/pci/msi.c

@@ -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.

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! Anyway, I've integrated all good suggestions from the
discussion into v3:
https://lore.kernel.org/lkml/20210825102636.52757-1-21cnbao@gmail.com/ (local)
So that we can have a new base to start after 5.14 is out.

Thanks,

        M.

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

Thanks
barry