On 08/26/2014 04:16 PM, David Gibson wrote: > On Fri, Aug 15, 2014 at 08:12:28PM +1000, Alexey Kardashevskiy wrote: >> Every sPAPRTCETable object holds an IOMMU memory region which holds >> a referenced to the sPAPRTCETable instance. So if we want to free >> an sPAPRTCETable instance, calling object_unref() will not be enough >> as embedded memory region will hold the reference and we need to break >> the loop. >> >> This adds a spapr_tce_free_table() helper which destroys the embedded >> memory region and then calls object_unref() on the sPAPRTCETable instance >> which will succeed now. The helper adds a new child under unique name >> derived from LIOBN. >> >> As we are here, fix spapr_tce_new_table() callers. >> At the moment spapr_tce_new_table() references sPAPRTCETable twice - >> once in object_new() and second time in object_property_add_child(). >> The callers of spapr_tce_new_table() should take care of correct >> referencing. > > So I've been trying to determine if there's any way to avoid this by > not constructing the reference loop in the first place, but all the > qom is breaking my brain.
Well. I could have added an additional object to sPAPRTCETable (just a pointer in the struct but not a QOM-child!) and use it as an owner for MemoryRegion. It would not hold reference to the table. But since I do not grasp the idea of having an owner for a MemoryRegion, this can break something which I cannot even imagine :) -- Alexey
