Re: [PATCH] xen/pciback: Make missing GSI non-fatal

[Jan Beulich <jbeulich@xxxxxxxx>]

On 26.02.2025 21:01, Jason Andryuk wrote:
> A PCI may not have a legacy IRQ.  In that case, do not fail assigning

Nit: Missing "device".

> to the pciback stub.  Instead just skip xen_pvh_setup_gsi().
> 
> This will leave psdev->gsi == -1.  In that case, when reading the value
> via IOCTL_PRIVCMD_PCIDEV_GET_GSI, return -ENOENT.  Userspace can used

Nit: "use".

> this to distinquish from other errors.

Nit: "distinguish".

> --- a/drivers/xen/acpi.c
> +++ b/drivers/xen/acpi.c
> @@ -101,7 +101,7 @@ int xen_acpi_get_gsi_info(struct pci_dev *dev,
>  
>       pin = dev->pin;
>       if (!pin)
> -             return -EINVAL;
> +             return -ENOENT;
>  
>       entry = acpi_pci_irq_lookup(dev, pin);
>       if (entry) {

While I can understand this change, ...

> @@ -116,7 +116,7 @@ int xen_acpi_get_gsi_info(struct pci_dev *dev,
>               gsi = -1;
>  
>       if (gsi < 0)
> -             return -EINVAL;
> +             return -ENOENT;
>  
>       *gsi_out = gsi;
>       *trigger_out = trigger;

... I'd expect this needs to keep using an error code other than ENOENT.
Aiui this path means the device has a pin-based interrupt, just that it's
not configured correctly. In which case we'd better not allow the device
to be handed to a guest. Unless there's logic in place (somewhere) to
make sure it then would get to see a device without pin-based interrupt.

> --- a/drivers/xen/xen-pciback/pci_stub.c
> +++ b/drivers/xen/xen-pciback/pci_stub.c
> @@ -240,6 +240,9 @@ static int pcistub_get_gsi_from_sbdf(unsigned int sbdf)
>       if (!psdev)
>               return -ENODEV;
>  
> +     if (psdev->gsi == -1)
> +             return -ENOENT;

This may, aiui, mean either of the two situations above. They would then
need distinguishing, too, if user space is intended to derive decisions
from the error code it gets.

Jan