Re: [RFC] kvm irq assignment

[Avi Kivity <avi@qumranet.com>]

Xu, Anthony wrote:
> Hi all,
> Thanks for your comments.
>
> I made this new patch based on your comments
>
> 1. use bimodal _PRT, to take advantage of IOAPIC pin 16~23
> 	the mapping is simple,  slot  ->  (slot&7)+16 IOAPIC pin,
> someone may provide good mapping ?
>   

I think it's fine. If we find a better one later, or if we add another 
ioapic, we can easily change it since the bios and qemu are shipped as a 
unit.

> 2. use ISA-bridge configure space 0x64 byte as a communication
> mechansim.
> 	When guest BIOS invokes _PIC, the value is passed to qemu
> through byte 0x64.
>              qemu know whether it is PIC mode and APIC mode by checking
> byte 0x64.
> 3. pci_slot_get_pirq and piix3_set_irq adopt different operation based
> on PIC mode/APIC mode
>   

I'm not sure how real hardware works, but I _think_ that it routes irqs 
unconditionally to both the legacy path and directly to the ioapic. So 
for example if slot 5 asserts an interrupt, we map it through the pci 
link mapping and generate an active high interrupt to one of {5, 10, 11} 
(both pic and ioapic), and simultaneously an active low interrupt to 
ioapic pin 21.

The _PIC method should disable the link interrupts if ioapic mode is 
disabled.

This removes the need for communication between the bios and qemu.

>  
> +            /* APIC and PIC flag */
> +            OperationRegion (P40D, PCI_Config, 0x64, 0x01)
> +
>   

This is actually SERIRQC, serial irq control.

> +
> +#ifdef KVM_CAP_IRQCHIP
>   

This should be unconditional.

> +static int pci_slot_get_pirq(PCIDevice *pci_dev, int irq_num)
> +{
> +    int slot_addend;
> +    if( piix3_dev->config[0x64])  // APIC mode
> +        return ((pci_dev->devfn >> 3) & 7)+16;
> +    else {		// PIC mode
> +        slot_addend = (pci_dev->devfn >> 3) - 1;
> +        return (irq_num + slot_addend) & 3;
> +    }
> +}
>   

What I'm suggesting is to "fork" the interrupt into two lines, one 
legacy path and the ioapic path.


-- 
I have a truly marvellous patch that fixes the bug which this
signature is too narrow to contain.