> > + /* > > + * Generations must be different for each address space. > > + * Init kvm generation close to the maximum to easily test the > > + * code of handling generation number wrap-around. > > + */ > > + slots->generation = i * 2 - 150; > > + rcu_assign_pointer(kvm->memslots[i], slots); > > } > > I can't seem to understand why rcu_assign_pointer wasn't used before. > kvm->memslots[i] was a rcu protected pointer even before this change, > right ? Actually, a better match is RCU_INIT_POINTER. Here there is no concurrent reader because we're just initializing the struct kvm. There is something else providing synchronization between this writer and the "first" RCU read-side. It could be signaling a condition variable, creating a thread, or releasing a mutex; all three of them have release semantics, which means they imply a smp_wmb just like rcu_assign_pointer does. Paolo > > if (init_srcu_struct(&kvm->srcu)) > > @@ -870,8 +872,14 @@ static struct kvm_memslots > > *install_new_memslots(struct kvm *kvm, > > * Increment the new memslot generation a second time. This prevents > > * vm exits that race with memslot updates from caching a memslot > > * generation that will (potentially) be valid forever. > > + * > > + * Generations must be unique even across address spaces. We do not need > > + * a global counter for that, instead the generation space is evenly > > split > > + * across address spaces. For example, with two address spaces, address > > + * space 0 will use generations 0, 4, 8, ... while * address space 1 will > > + * use generations 2, 6, 10, 14, ... > > */ > > - slots->generation++; > > + slots->generation += KVM_ADDRESS_SPACE_NUM * 2 - 1; > > > > kvm_arch_memslots_updated(kvm, slots); >
