On Wed, May 17, 2023 at 11:32 PM Joao Martins <joao.m.martins@xxxxxxxxxx> wrote: > > On 17/05/2023 13:04, dengqiao.joey wrote: > > This seems to be a different issue. Joao's patch tries to fix the issue > > that IRTE is not changed. The problem I encountered is that IRTE does > > get changed, thus the destination field is also cleared by > > amd_iommu_activate_guest_mode(). > > Whether the amd_iommu_activate_guest_mode() clears the destination field or not > doesn't change the situation I think. So I don't think that is the problem > you're having. amd_iommu_activate_guest_mode() will always make IRTEs with > isRun=0. Which means your VF interrupts get injected via GALog events and the > DestID is meaningless there :/ Even if you leave the old affinity there with > isRun=1 bit it's still wrong as by the time you run the vcpu, the wrong affinity > will be used by the VF. More fundamentally I am not sure that on IRTE routing > updates in the VM that you can look on vcpu->cpu the way you are using it could > be potentially invalid when vcpus are blocked/preempted. You're changing IRTEs > in the whole VM in pi_update_irte() > > My changes essentially handle the fact where IRTEs will be always updated with > the right GATag set in the IRTE ... and then when the vCPU migrates, wakes or > block+unblocks it will change IRTE[DestID] with the new pcpu (which the vcpu is > running on) via amd_iommu_update_ga(). So there's a good chance my changes fixes > your affinity issue. > > Joao Sorry I forgot to mention before that in my application scenario, the vcpu runs on a dedicated isolated cpu and uses "mwait" instead of "halt". So it will not be migrated, blocked/wake and rarely get preempted by other threads. I think your patch can not fix my issue because the vcpu rarely gets the opportunity to execute amd_iommu_update_ga() from vcpu_load(). So each time the interrupt affinity is updated I can observe the loss of VF interrupt. After applying my patch, the problem is solved.
