[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] Re: [Xen-devel] [PATCH v3 32/39] ARM: new VGIC: Implement arch_move_irqs()
On Wed, 21 Mar 2018, Andre Przywara wrote: > When a VCPU moves to another CPU, we need to adjust the target affinity > of any hardware mapped vIRQs, to observe our "physical-follows-virtual" > policy. > Implement arch_move_irqs() to adjust the physical affinity of all hardware > mapped vIRQs targetting this VCPU. > > Signed-off-by: Andre Przywara <andre.przywara@xxxxxxxxxx> > Reviewed-by: Julien Grall <julien.grall@xxxxxxx> > --- > xen/arch/arm/vgic/vgic.c | 39 +++++++++++++++++++++++++++++++++++++++ > 1 file changed, 39 insertions(+) > > diff --git a/xen/arch/arm/vgic/vgic.c b/xen/arch/arm/vgic/vgic.c > index ffab0b2635..23b8abfc5e 100644 > --- a/xen/arch/arm/vgic/vgic.c > +++ b/xen/arch/arm/vgic/vgic.c > @@ -791,6 +791,45 @@ void gic_dump_vgic_info(struct vcpu *v) > spin_unlock_irqrestore(&v->arch.vgic.ap_list_lock, flags); > } > > +/** > + * arch_move_irqs() - migrate the physical affinity of hardware mapped vIRQs > + * @v: the vCPU, already assigned to the new pCPU > + * > + * arch_move_irqs() updates the physical affinity of all virtual IRQs > + * targetting this given vCPU. This only affects hardware mapped IRQs. The > + * new pCPU to target is already set in v->processor. > + * This is called by the core code after a vCPU has been migrated to a new > + * physical CPU. > + */ > +void arch_move_irqs(struct vcpu *v) > +{ > + struct domain *d = v->domain; > + unsigned int i; > + > + /* We only target SPIs with this function */ > + for ( i = 0; i < d->arch.vgic.nr_spis; i++ ) > + { > + struct vgic_irq *irq = vgic_get_irq(d, NULL, i + > VGIC_NR_PRIVATE_IRQS); > + unsigned long flags; > + > + if ( !irq ) > + continue; > + > + spin_lock_irqsave(&irq->irq_lock, flags); > + > + /* only vIRQs that are not on a vCPU yet , but targetting this vCPU > */ > + if ( irq->hw && !irq->vcpu && irq->target_vcpu == v) > + { In vgic_mmio_write_target, we change the physical irq affinity immediately, without checking for !irq->vcpu. I think it is OK because if a second interrupt for vcpuB comes in cpuB while it is still injected in vcpuA/cpuA, vgic_get_irq returns the same vgic_irq instance, vgic_inject_irq sets pending_latch to true. vgic_queue_irq_unlock does nothing because irq->vcpu is set. Then when vcpuA traps into Xen, vgic_prune_ap_list will take care of moving the vgic_irq to the ap_list belonging to vcpuB. This seems to work, but don't we also need a vcpu_kick at the end of vgic_prune_ap_list to make sure the changes take effect in vcpuB? vcpuB could take an very long time to trap back into Xen again. But the real question is: why do we need to check for !irq->vcpu here? And worse: if an interrupt has irq->vcpu set, then who will take care of fixing the physical irq affinity later? It looks like we should remove the "&& !irq->vcpu" here so that we can rely on the same mechanism already in place for ITARGETSR changes. However, would that work with already active interrupts? I think it should but I wanted to double check. > + irq_desc_t *desc = irq_to_desc(irq->hwintid); > + > + irq_set_affinity(desc, cpumask_of(v->processor)); > + } > + > + spin_unlock_irqrestore(&irq->irq_lock, flags); > + vgic_put_irq(d, irq); > + } > +} > + > struct irq_desc *vgic_get_hw_irq_desc(struct domain *d, struct vcpu *v, > unsigned int virq) > { _______________________________________________ Xen-devel mailing list Xen-devel@xxxxxxxxxxxxxxxxxxxx https://lists.xenproject.org/mailman/listinfo/xen-devel
|
Lists.xenproject.org is hosted with RackSpace, monitoring our |