----- Original Message -----
> From: "Alex Williamson" <alex.william...@redhat.com>
> To: "Timothy Pearson" <tpear...@raptorengineering.com>
> Cc: "kvm" <k...@vger.kernel.org>, "linuxppc-dev"
> <linuxppc-dev@lists.ozlabs.org>
> Sent: Monday, September 22, 2025 1:01:22 PM
> Subject: Re: [PATCH v2] vfio/pci: Fix INTx handling on legacy non-PCI 2.3
> devices
> On Mon, 22 Sep 2025 12:18:34 -0500 (CDT)
> Timothy Pearson <tpear...@raptorengineering.com> wrote:
>
>> PCI devices prior to PCI 2.3 both use level interrupts and do not support
>> interrupt masking, leading to a failure when passed through to a KVM guest on
>> at least the ppc64 platform. This failure manifests as receiving and
>> acknowledging a single interrupt in the guest, while the device continues to
>> assert the level interrupt indicating a need for further servicing.
>>
>> When lazy IRQ masking is used on DisINTx- (non-PCI 2.3) hardware, the
>> following
>> sequence occurs:
>>
>> * Level IRQ assertion on device
>> * IRQ marked disabled in kernel
>> * Host interrupt handler exits without clearing the interrupt on the device
>> * Eventfd is delivered to userspace
>> * Host interrupt controller sees still-active INTX, reasserts IRQ
>> * Host kernel ignores disabled IRQ
>> * Guest processes IRQ and clears device interrupt
>> * Software mask removed by VFIO driver
>
> This isn't the sequence that was previously identified as the issue.
> An interrupt controller that reasserts the IRQ when it remains active
> is not susceptible to the issue, and is what I think we generally
> expect on x86. I understand that we believe this issue manifests
> exactly because the interrupt controller does not reassert an interrupt
> that remains active. I think the sequence is:
>
> * device asserts INTx
> * vfio_intx_handler() calls disable_irq_nosync() to mark IRQ disabled
> * interrupt delivered to userspace via eventfd
> * userspace driver/VM services interrupt
> * device de-asserts INTx
> * device re-asserts INTx
> * interrupt received while IRQ disabled is masked at controller
> * VMM performs EOI via unmask ioctl, enable_irq() clears disable and
> unmasks IRQ
> * interrupt controller does not reassert interrupt to the host
>
> The fix then, aiui, is that disabling the unlazy mode masks the
> interrupt at disable_irq_nosync(), the same sequence of de-asserting
> and re-asserting the interrupt occurs at the controller, but since the
> controller was masked at the new rising edge, it will send the
> interrupt when umasked.
That is correct. Technically we're dealing with two different ways to hang the
controller, but this one is the most likely; both fundamentally boil down to
not receiving a new INTx falling edge (INTx is active low) after the interrupt
is unmasked. I've updated the description to match.
>> The behavior is now platform-dependent. Some platforms (amd64) will continue
>> to spew IRQs for as long as the INTX line remains asserted, therefore the IRQ
>> will be handled by the host as soon as the mask is dropped. Others (ppc64)
>> will
>> only send the one request, and if it is not handled no further interrupts
>> will
>> be sent. The former behavior theoretically leaves the system vulnerable to
>> interrupt storm, and the latter will result in the device stalling after
>> receiving exactly one interrupt in the guest.
>>
>> Work around this by disabling lazy IRQ masking for DisINTx- INTx devices.
>> ---
>> drivers/vfio/pci/vfio_pci_intrs.c | 5 +++++
>> 1 file changed, 5 insertions(+)
>>
>> diff --git a/drivers/vfio/pci/vfio_pci_intrs.c
>> b/drivers/vfio/pci/vfio_pci_intrs.c
>> index 123298a4dc8f..d8637b53d051 100644
>> --- a/drivers/vfio/pci/vfio_pci_intrs.c
>> +++ b/drivers/vfio/pci/vfio_pci_intrs.c
>> @@ -304,6 +304,9 @@ static int vfio_intx_enable(struct vfio_pci_core_device
>> *vdev,
>>
>> vdev->irq_type = VFIO_PCI_INTX_IRQ_INDEX;
>>
>> + if (!vdev->pci_2_3)
>> + irq_set_status_flags(pdev->irq, IRQ_DISABLE_UNLAZY);
>> +
>> ret = request_irq(pdev->irq, vfio_intx_handler,
>> irqflags, ctx->name, ctx);
>> if (ret) {
>
> This branch is an example of where we're not clearing the flag on
> error. Thanks,
Whoops! That's what I get for not looking closely and grepping for free_irq()!
V3 sent.
