Re: Xen regression, Was: [PATCH] x86/irq: Probe for PIC presence before allocating descs for legacy IRQs

[David Vrabel <david.vrabel@citrix.com>]

On 12/04/16 19:06, Stefano Stabellini wrote:
> On Tue, 12 Apr 2016, Boris Ostrovsky wrote:
>> On 04/11/2016 10:08 PM, Stefano Stabellini wrote:
>>> Hi all,
>>>
>>> Unfortunately this patch (now commit
>>> 8c058b0b9c34d8c8d7912880956543769323e2d8) causes a regression on Xen
>>> when running on top of QEMU: the number of PIT irqs get set to 0 by
>>> probe_8259A but actually there are 16.
>>>
>>> Any suggestions on how to fix this?
>>>
>>> 1) we could revert 8c058b0b9c34d8c8d7912880956543769323e2d8
>>> 2) we could introduce an 'if (!xen_domain())' in probe_8259A
>>> 3) suggestions welcome
>>
>> Stefano, do you have b4ff8389ed14b849354b59ce9b360bdefcdbf99c ?
>>
>> It was supposed to fix this problem for Xen. However, I just noticed that
>> arch/arm64/include/asm/irq.h makes nr_legacy_irqs() return 0 (unlike
>> arch/arm/include/asm/irq.h). Could that be the problem?
> 
> I have b4ff8389ed14b849354b59ce9b360bdefcdbf99c but it doesn't fix the
> issue for me.
> 
> Is the idea of your patch that xen_allocate_irq_gsi will allocate the
> descriptor dynamically instead?  If so, it doesn't work because it
> doesn't get called for irq 14:
> 
> piix_init_one -> ata_pci_sff_activate_host -> devm_request_irq ->
> devm_request_threaded_irq-> request_threaded_irq -> irq_to_desc(14) ->
> -EVAIL
> 
> If you look at pci_xen_initial_domain, the loop:
> 
> for (irq = 0; irq < nr_legacy_irqs(); irq++) {
> 
> won't work anymore because by the time is called, nr_legacy_irqs()
> already returns 0, because it has been changed by probe_8259A().
> 
> We also need the following patch:
> 
> ---
> 
> xen/x86: actually allocate legacy interrupts on PV guests
> 
> b4ff8389ed14 is incomplete: relies on nr_legacy_irqs() to get the number
> of legacy interrupts when actually nr_legacy_irqs() returns 0 after
> probe_8259A(). Use NR_IRQS_LEGACY instead.
> 
> Signed-off-by: Stefano Stabellini <sstabellini@kernel.org>
> 
> diff --git a/arch/x86/pci/xen.c b/arch/x86/pci/xen.c
> index beac4df..6db0060 100644
> --- a/arch/x86/pci/xen.c
> +++ b/arch/x86/pci/xen.c
> @@ -492,7 +492,7 @@ int __init pci_xen_initial_domain(void)
>  	__acpi_register_gsi = acpi_register_gsi_xen;
>  	__acpi_unregister_gsi = NULL;
>  	/* Pre-allocate legacy irqs */
> -	for (irq = 0; irq < nr_legacy_irqs(); irq++) {
> +	for (irq = 0; irq < NR_IRQS_LEGACY; irq++) {
>  		int trigger, polarity;
>  
>  		if (acpi_get_override_irq(irq, &trigger, &polarity) == -1)

I think I prefer this fix because PV guests don't have a legacy PIC and
don't need any legacy irqs allocated so fiddling with nr_legacy_irqs()
seems wrong.

But we do need a comment here saying something like:

/*
 * Pre-allocate the legacy IRQs.  Use NR_LEGACY_IRQS here
 * because we don't have a PCI and thus nr_legacy_irqs() is zero.
 */

Does this make sense?

David