Messages in this thread | | | Subject | Re: [PATCH] kvm: lock slots_lock around device assignment | From | Alex Williamson <> | Date | Wed, 18 Apr 2012 20:58:24 -0600 |
| |
On Wed, 2012-04-18 at 23:30 -0300, Marcelo Tosatti wrote: > On Tue, Apr 17, 2012 at 09:46:44PM -0600, Alex Williamson wrote: > > @@ -340,7 +343,11 @@ int kvm_iommu_unmap_guest(struct kvm *kvm) > > if (!domain) > > return 0; > > > > + mutex_lock(&kvm->slots_lock); > > kvm_iommu_unmap_memslots(kvm); > > + kvm->arch.iommu_domain = NULL; > > + mutex_unlock(&kvm->slots_lock); > > + > > iommu_domain_free(domain); > > return 0; > > } > > This might trigger lockdep warnings due to > > kvm_vm_ioctl_create_vcpu > mutex_lock(&kvm->lock) > kvm_put_kvm(kvm) > kvm_destroy_vm > kvm_iommu_unmap_guest > > sequence. > > Better drop it, it is not necessary in vm destruction > path (since only user is self).
I actually ran this with lockdep and didn't generate a warning; hopefully I had it configured correctly. Also, we'll soon be unmapping the guest any time we remove the last assigned device so this will no longer be a vm destruction-only path. We can just as easily race adding new mappings or removing already removed ones on that path. We also acquire kvm->lock in the mapping path:
kvm_vm_ioctl_assign_device() { mutex_lock(&kvm->lock); if (!kvm->arch.iommu_domain) { r = kvm_iommu_map_guest(kvm);
which by inspection and the lock ordering note in kvm_main seems to be ok. Thanks,
Alex
| |