Re: [PATCH] PCI: vmd: Remove MSI-X check on child devices

[Bjorn Helgaas <helgaas@kernel.org>]

On Tue, Aug 12, 2025 at 08:49:50AM -0600, Keith Busch wrote:
> On Tue, Aug 12, 2025 at 08:27:15AM +0200, Nam Cao wrote:
> > On Mon, Aug 11, 2025 at 05:46:59PM -0500, Bjorn Helgaas wrote:
> > > On Mon, Aug 11, 2025 at 07:39:35AM +0200, Nam Cao wrote:
> > > > Commit d7d8ab87e3e7 ("PCI: vmd: Switch to msi_create_parent_irq_domain()")
> > > > added a WARN_ON sanity check that child devices support MSI-X, because VMD
> > > > document says [1]:
> > > > 
> > > >     "Intel VMD only supports MSIx Interrupts from child devices and
> > > >     therefore the BIOS must enable PCIe Hot Plug and MSIx interrups [sic]."
> > > 
> > > Can VMD tell the difference between an incoming MSI MWr transaction
> > > and an MSI-X MWr?
> > > 
> > > I wonder if "MSIx" was meant to mean "VMD only supports MSI or MSI-X
> > > interrupts, not INTx interrupts, from child devices"?
> > 
> > I don't know, sorry. I am hoping that the VMD maintainers can help us here.
> 
> The doc you linked is riddled with errors. The original vmd commit
> message is more accurate: VMD domains support child devices with MSI and
> MSI-x interrupts. The VMD device can't even tell the difference which
> one the device is using. It just manipulates messages sent to the usual
> APIC address 0xfeeXXXXX.

Thanks, Keith!  I updated the commit log like this:

  d7d8ab87e3e7 ("PCI: vmd: Switch to msi_create_parent_irq_domain()") added a
  WARN_ON sanity check that child devices support MSI-X, because VMD document
  says [1]:

    Intel VMD only supports MSIx Interrupts from child devices and therefore
    the BIOS must enable PCIe Hot Plug and MSIx interrups [sic].

  However, the VMD device can't even tell the difference between a child
  device using MSI and one using MSI-X.  Per 185a383ada2e ("x86/PCI: Add
  driver for Intel Volume Management Device (VMD)"), VMD does not support
  INTx interrupts, but does support child devices using either MSI or MSI-X.

  Remove the sanity check to avoid the unnecessary WARN_ON reported by Ammar.