On Thu, Aug 07, 2025 at 07:43:17PM +0200, Nam Cao wrote:
> Bjorn Helgaas <helgaas@xxxxxxxxxx> writes:
> > On Thu, Aug 07, 2025 at 10:10:51AM +0200, Nam Cao wrote:
> >> vmd_msi_alloc() allocates struct vmd_irq and stashes it into
> >> irq_data->chip_data associated with the VMD's interrupt domain.
> >> vmd_msi_free() extracts the pointer by calling irq_get_chip_data() and
> >> frees it.
> >>
> >> irq_get_chip_data() returns the chip_data associated with the top interrupt
> >> domain. This worked in the past, because VMD's interrupt domain was the top
> >> domain.
> >>
> >> But since commit d7d8ab87e3e7 ("PCI: vmd: Switch to
> >> msi_create_parent_irq_domain()") changed the interrupt domain hierarchy,
> >> VMD's interrupt domain is not the top domain anymore. irq_get_chip_data()
> >> now returns the chip_data at the MSI devices' interrupt domains. It is
> >> therefore broken for vmd_msi_free() to kfree() this chip_data.
> >>
> >> Fix this issue, correctly extract the chip_data associated with the VMD's
> >> interrupt domain.
> ...
> >
> > Applied to pci/for-linus for v6.17, thanks!
> >
> > I assume you checked the other msi_create_parent_irq_domain() changes
> > for similar problems?
>
> Not before you reminded me :(
>
> But yes, none of the similar PCI patches has the same problem.
Great, thanks for checking! I'll try to get this in before v6.17-rc1
since many people will start using that.
Bjorn
