On Tue, Nov 10, 2020 at 12:46:48PM -0500, Eduardo Habkost wrote: > On Tue, Nov 10, 2020 at 04:08:16PM +0100, Paolo Bonzini wrote: > > On 10/11/20 16:03, Eduardo Habkost wrote: > > > > Does anyone have any arguments for which solution is preferred? > > > I'd say (2) is preferred, as we don't expect object_new(T) to > > > have any side effects outside the object instance state. > > > > Since there are no listeners, the side effects of address_space_init() are > > relatively limited. So doing it in instance_init is not a big deal. > > > > > Most > > > address_space_init() calls in the code today seem to be in > > > realize functions. > > > > > > However, I wonder if we could make this simpler (and mistakes > > > less fatal) if we make AddressSpace a QOM child of the device. > > > Paolo, would it be too much overhead to make AddressSpace a QOM > > > object? > > > > No, it wouldn't. AddressSpace is already quite heavyweight. > > I thought this was going to be an easy job, but call_rcu() > requires rcu_head to be the first struct field. I assume it is > acceptable to use call_rcu1() + container_of() manually in this > case.
Wait. What exactly prevents callers of address_space_destroy() from freeing the area containing the AddressSpace struct before do_address_space_destroy() gets a chance to be called? -- Eduardo
