(I think the Cc list is too long here, but then I don't know who to
keep and who to possibly drop.)
On 12.12.2023 09:49, Roger Pau Monné wrote:
> On Tue, Dec 12, 2023 at 06:16:43AM +0000, Chen, Jiqian wrote:
>> On 2023/12/11 23:45, Roger Pau Monné wrote:
>>> On Wed, Dec 06, 2023 at 06:07:26AM +0000, Chen, Jiqian wrote:
>>>> +static int xen_pvh_setup_gsi(gsi_info_t *gsi_info)
>>>> +{
>>>> + struct physdev_setup_gsi setup_gsi;
>>>> +
>>>> + setup_gsi.gsi = gsi_info->gsi;
>>>> + setup_gsi.triggering = (gsi_info->trigger == ACPI_EDGE_SENSITIVE ? 0 : 1);
>>>> + setup_gsi.polarity = (gsi_info->polarity == ACPI_ACTIVE_HIGH ? 0 : 1);
>>>> +
>>>> + return HYPERVISOR_physdev_op(PHYSDEVOP_setup_gsi, &setup_gsi);
>>>> +}
>>>
>>> Hm, why not simply call pcibios_enable_device() from pciback? What
>> pcibios_enable_device had been called when using cmd "xl pci-assignable-add sbdf" from pciback. But it didn't do map_pirq and setup_gsi.
>> Because pcibios_enable_device-> pcibios_enable_irq-> __acpi_register_gsi(acpi_register_gsi_ioapic PVH specific)
>>> you are doing here using the hypercalls is a backdoor into what's done
>>> automatically by Xen on IO-APIC accesses by a PVH dom0.
>> But the gsi didn't be unmasked, and vioapic_hwdom_map_gsi is never called.
>> So, I think in pciback, if we can do what vioapic_hwdom_map_gsi does.
>>
>
> I see, it does setup the IO-APIC pin but doesn't unmask it, that's
> what I feared.
>
>>> It will be much more natural for the PVH dom0 model to simply use the
>>> native way to configure and unmask the IO-APIC pin, and that would
>>> correctly setup the triggering/polarity and bind it to dom0 without
>>> requiring the usage of any hypercalls.
>> Do you still prefer that I called unmask_irq in pcistub_init_device, as this v2 patch do?
>> But Thomas Gleixner think it is not suitable to export unmask_irq.
>
> Yeah, that wasn't good.
>
>>>
>>> Is that an issue since in that case the gsi will get mapped and bound
>>> to dom0?
>> Dom0 do map_pirq is to pass the check xc_domain_irq_permission()-> pirq_access_permitted(),
>
> Can we see about finding another way to fix this check?
>
> One option would be granting permissions over the IRQ in
> PHYSDEVOP_setup_gsi?
There's no domain available there, and imo it's also the wrong interface to
possibly grant any permissions.
> Otherwise we could see about modifying the logic in PHYSDEVOP_map_pirq
> so that the hardware domain can map IRQs to other domains without
> having it mapped to itself first?
While this may be possible to arrange for, it still would feel wrong. How
would you express the same in a disaggregated environment? I.e. how would
you make sure a domain trying to grant permission is actually permitted to
do so for the resource in question?
> I think the call to PHYSDEVOP_setup_gsi in pciback is fine, but I
> would really like to avoid the usage of PHYSDEVOP_{,un}map_pirq on a
> PVH dom0 against itself.
+1
Jan
