On 29/01/19 10:32, Kang, Luwei wrote:
>>> However, you should at least change the comment in vcpu_enter_guest to
>>> mention "before reading PIR" instead of "before reading PIR.ON".
>>
>> Will do that. I think the "checking PIR.ON" should be PID.ON. I will fix it.
>>
> Hi Paolo,
> I reconsidered the comment in vcpu_enter_guest() and think it may don't need to change. The original comment as below:
> * 2) For APICv, we should set ->mode before checking PIR.ON. This
> * pairs with the memory barrier implicit in pi_test_and_set_on
> * (see vmx_deliver_posted_interrupt).
>
> I think "before checking PIR.ON" is mean "before checking PID.PIR and PID.ON".
I can say it only means PID.ON because I wrote the comment. :)
The idea is that checking ON is enough: KVM assumes that PID.PIR is only
set if PID.ON is set, because it follows the definition of ON in table
29-1 of the SDM: "If this bit is set, there is a notification
outstanding for one or more posted interrupts in bits 255:0".
VT-D breaks this assumption whenever SN=1 ("hardware does not generate
notification event nor modify the ON field"), resulting in nonzero
PID.PIR but PID.ON=0. I'm sure there was a reason for that, but it does
result in inconsistency between the PID definitions in the SDM and the
VT-D specification. The right fix is definitely to reconcile this
difference and test the bitmap after SN is cleared (with
smp_mb__after_atomic after clearing SN), and set ON=1 if the bitmap is
not clear.
Paolo
