Re: [XEN PATCH v13 3/6] x86/pvh: Add PHYSDEVOP_setup_gsi for PVH dom0

[Jan Beulich <jbeulich@xxxxxxxx>]

On 16.08.2024 13:08, Jiqian Chen wrote:
> The gsi of a passthrough device must be configured for it to be
> able to be mapped into a hvm domU.
> But When dom0 is PVH, the gsis may not get registered(see below
> clarification), it causes the info of apic, pin and irq not be
> added into irq_2_pin list, and the handler of irq_desc is not set,
> then when passthrough a device, setting ioapic affinity and vector
> will fail.
> 
> To fix above problem, on Linux kernel side, a new code will
> need to call PHYSDEVOP_setup_gsi for passthrough devices to
> register gsi when dom0 is PVH.
> 
> So, add PHYSDEVOP_setup_gsi into hvm_physdev_op for above
> purpose.
> 
> Clarify two questions:
> First, why the gsi of devices belong to PVH dom0 can work?
> Because when probe a driver to a normal device, it uses the normal
> probe function of pci device, in its callstack, it requests irq
> and unmask corresponding ioapic of gsi, then trap into xen and
> register gsi finally.
> Callstack is(on linux kernel side) pci_device_probe->
> request_threaded_irq-> irq_startup-> __unmask_ioapic->
> io_apic_write, then trap into xen hvmemul_do_io->
> hvm_io_intercept-> hvm_process_io_intercept->
> vioapic_write_indirect-> vioapic_hwdom_map_gsi-> mp_register_gsi.
> So that the gsi can be registered.
> 
> Second, why the gsi of passthrough device can't work when dom0
> is PVH?
> Because when assign a device to passthrough, it uses the specific
> probe function of pciback, in its callstack, it doesn't install a
> fake irq handler due to the ISR is not running. So that
> mp_register_gsi on Xen side is never called, then the gsi is not
> registered.
> Callstack is(on linux kernel side) pcistub_probe->pcistub_seize->
> pcistub_init_device-> xen_pcibk_reset_device->
> xen_pcibk_control_isr->isr_on==0.

So: Underlying XSA-461 was the observation that the very limited set of
cases where this fake IRQ handler is installed is an issue. The problem
of dealing with "false" IRQs when a line-based interrupt is shared
between devices affects all parties, not just Dom0 and not just PV
guests. Therefore an alternative to the introduction of a new hypercall
would be to simply leverage that the installation of such a handler
will need widening anyway.

However, the installation of said handler presently also occurs in
cases where it's not really needed - when the line isn't shared. Thus,
if the handler registration would also be eliminated when it's not
really needed, we'd be back to needing a separate hypercall.

So I think first of all it needs deciding what is going to be done in
Linux, at least in pciback (as here we care about the Dom0 case only).

Jan