Re: [PATCH] x86/dpci: remove the dpci EOI timer

[Jan Beulich <jbeulich@xxxxxxxx>]

On 14.01.2021 12:56, Andrew Cooper wrote:
> On 14/01/2021 11:48, Jan Beulich wrote:
>> On 13.01.2021 14:11, Roger Pau Monné wrote:
>>> On Wed, Jan 13, 2021 at 06:21:03AM +0000, Tian, Kevin wrote:
>>>>> From: Roger Pau Monne <roger.pau@xxxxxxxxxx>
>>>>> As with previous patches, I'm having a hard time figuring out why this
>>>>> was required in the first place. I see no reason for Xen to be
>>>>> deasserting the guest virtual line. There's a comment:
>>>>>
>>>>> /*
>>>>>  * Set a timer to see if the guest can finish the interrupt or not. For
>>>>>  * example, the guest OS may unmask the PIC during boot, before the
>>>>>  * guest driver is loaded. hvm_pci_intx_assert() may succeed, but the
>>>>>  * guest will never deal with the irq, then the physical interrupt line
>>>>>  * will never be deasserted.
>>>>>  */
>>>>>
>>>>> Did this happen because the device was passed through in a bogus state
>>>>> where it would generate interrupts without the guest requesting
>>>> It could be a case where two devices share the same interrupt line and
>>>> are assigned to different domains. In this case, the interrupt activity of 
>>>> two devices interfere with each other.
>>> This would also seem to be problematic if the device decides to use
>>> MSI instead of INTx, but due to the shared nature of the INTx line we
>>> would continue to inject INTx (generated from another device not
>>> passed through to the guest) to the guest even if the device has
>>> switched to MSI.
>> I'm having trouble with this: How does the INTx line matter when
>> a device is using MSI? I don't see why we should inject INTx when
>> the guest has configured a device for MSI; if we do, this would
>> indeed look like a bug (but aiui we bind either the MSI IRQ or
>> the pin based one, but not both).
> 
> When MSI is configured, a spec-complient device shouldn't raise INTx. 
> But there are plenty of quirks in practice, and ample opportunity for
> races, considering that in a Xen system, there are at least 3 entities
> in the system fighting over control of the device.
> 
> I suspect the problem is "what happens when Xen gets an INTx".  We don't
> know which device it came from, and therefore we can't even attempt to
> filter out the devices we suspect are using MSI, in an attempt to avoid
> raising the line with every domain.

The domain using MSI won't have the INTx IRQ bound, so won't be
notified of the INTx instance at all. What is likely to happen
is that the other domain(s) sharing the same INTx IRQ would get
notified, but won't find anything to do. Which in turn may lead
to the EOI timer getting engaged, and hence IRQs over this line
getting throttled.

Jan