RE: Virtualizing MSI-X on IMS via VFIO

[Thomas Gleixner <tglx@linutronix.de>]

On Fri, Jun 25 2021 at 10:43, Thomas Gleixner wrote:
> On Fri, Jun 25 2021 at 05:21, Kevin Tian wrote:
>> p.s. one question to Thomas. As Alex cited above, software must 
>> not modify the Address, Data, or Steering Tag fields of an MSI-X
>> entry while it is unmasked. However this rule might be violated
>> today in below flow:
>>
>> request_irq()
>>     __setup_irq()
>>         irq_startup()
>>             __irq_startup()
>>                 irq_enable()
>>                     unmask_irq() <<<<<<<<<<<<<
>>         irq_setup_affinity()
>>             irq_do_set_affinity()
>>                 msi_set_affinity() // when IR is disabled
>>                     irq_msi_update_msg()
>>                         pci_msi_domain_write_msg() <<<<<<<<<<<<<<
>>
>> Isn't above have msi-x entry updated after it's unmasked? 
>
> Dammit, I could swear that we had masking at the core or PCI level at
> some point. Let me dig into this.

Indeed, that code path does not check irq_can_move_pcntxt(). It doesn't
blow up in our face by chance because of this:

     __setup_irq()
        irq_activate()
        unmask()
        irq_setup_affinity()

irq_activate() assigns a vector based on the affinity mask so
irq_setup_affinity() ends up writing the same data again pointlessly.

For some stupid reason the ordering of startup/setup_affinity is the way
it is for historical reasons. I tried to reorder it at some point but
that caused failure on !x86 so I went back to the status quo.

All other affinity settings happen with the interrupt masked because we
do that from actual interrupt context via irq_move_masked_irq() which
does the right thing.

Let me fix that proper for the startup case.

Thanks,

        tglx