On Sat, Jul 28, 2018 at 11:07:33AM +0200, Mike Galbraith wrote:

b103cc3f10c0 ("KVM: arm/arm64: Avoid timer save/restore in vcpu entry/exit")
requires vgic_irq.irq_lock be converted to raw_spinlock_t.

Problem: kvm_preempt_ops.sched_in = kvm_sched_in;
   kvm_sched_in()
      kvm_arch_vcpu_load()
         kvm_timer_vcpu_load() <- b103cc3f10c0 addition
            kvm_timer_vcpu_load_gic()
               kvm_vgic_map_is_active()
                  spin_lock_irqsave(&irq->irq_lock, flags);

Quoting virt/kvm/arm/vgic/vgic.c, locking order is...

  kvm->lock (mutex)
    its->cmd_lock (mutex)
      its->its_lock (mutex)
        vgic_cpu->ap_list_lock         must be taken with IRQs disabled
          kvm->lpi_list_lock           must be taken with IRQs disabled
            vgic_irq->irq_lock         must be taken with IRQs disabled

...meaning vgic_dist.lpi_list_lock and vgic_cpu.ap_list_lock must be
converted as well.

The thing missing from the Changelog is the analysis that all the work
done under these locks is indeed properly bounded and cannot cause
excessive latencies.