Re: [Xen-devel] [PATCH for-4.13 v3] x86/passthrough: fix migration of MSI when using posted interrupts

[Joe Jin <joe.jin@xxxxxxxxxx>]

On 11/15/19 6:06 AM, Roger Pau Monné wrote:
> On Fri, Nov 15, 2019 at 05:23:51AM +0000, Tian, Kevin wrote:
>>> From: Roger Pau Monne [mailto:roger.pau@xxxxxxxxxx]
>>> Sent: Friday, November 8, 2019 9:34 PM
>>>
>>> When using posted interrupts and the guest migrates MSI from vCPUs Xen
>>> needs to flush any pending PIRR vectors on the previous vCPU, or else
>>> those vectors could get wrongly injected at a later point when the MSI
>>> fields are already updated.
>> I may overlook but isn't it the guest's responsibility of handling such
>> case? Even on bare metal, an in-fly interrupt may be delivered to
>> wrong CPU when MSI is being migrated?
> According to Joe from Oracle Linux already takes care of that by
> checking IRR when migrating interrupts between CPUs, but it seems like
> the vector is not pending in IRR (my hypothesis is that it's pending
> in PIR but lacking a sync into IRR).
> 
> After digging more into the posted interrupt code, I think there's an
> issue somewhere else, and the sync on MSI reconfiguration done by this
> patch is just covering that up.
> 
> There shouldn't be any interrupts pending in the PIR when the vCPU is
> running, and any pending vectors in the PIR should be synced into IRR
> before executing the vCPU.
> 
> AFAICT there's an issue with how PIR is synced into IRR, it relies on
> vlapic_find_highest_irr being called from vlapic_has_pending_irq, but
> depending on which interrupts are pending it's possible that
> vlapic_has_pending_irq is not called by hvm_vcpu_has_pending_irq, thus
> leaving IRR stale.
> 
> The patch below should solve that and also simplify
> __vmx_deliver_posted_interrupt, there's no reason to raise a softirq
> in __vmx_deliver_posted_interrupt: if the vCPU is the one currently
> running or if it's not running at all the sync of PIR to IRR will
> happen on vmentry, without the need of any softirq being set. Also
> note the raise_softirq in __vmx_deliver_posted_interrupt should have
> been a cpu_raise_softirq(cpu, VCPU_KICK_SOFTIRQ) instead.
> 
> Joe, can you give a try to the patch below?

This patch fixed my issue as well.

Thanks,
Joe

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel