On Mon, May 13, 2013 at 02:19:50PM +0200, Paolo Bonzini wrote: > Il 13/05/2013 14:07, Peter Maydell ha scritto: > > On 13 May 2013 12:48, David Gibson <da...@gibson.dropbear.id.au> wrote: > >> On Mon, May 13, 2013 at 12:10:10PM +0100, Peter Maydell wrote: > >>> Hmm, is this the ideal semantics? Typically the owner of the > >>> MemoryListener isn't the owner of the AddressSpace so it isn't > >>> necessarily in a position to guarantee that it can unregister > >>> the listener before the address space is destroyed. In fact > >>> as the listener API is currently documented, the filter > >>> argument is just an optimisation to save the callbacks having > >>> to filter out irrelevant information themselves. > >> > >> If so, then it's broken by design. There's no guarantee that after an > >> AddressSpace is destroyed another one won't be created at the same > >> address (in fact, depending on your malloc() implementation, it could > >> be very likely). So references by pointer to an object *must* be > >> removed before the object itself is freed. > > > > Mmm. Looking through the code it turns out we don't actually > > make use of the ability to pass NULL as a filter (except in > > target-arm/kvm.c which was just me being lazy and not passing > > in the system address space). Perhaps we should just drop that > > capability, at which point you have a clearer "you are listening > > on one AS and you must make sure you arrange to unregister before > > that AS goes away" API definition? > > Yes, that could be an idea.
Fine by me, that would also naturally lead to making the listener list per-AS. Either way, it doesn't alter the fact that pointer references to the AS need to be removed before the AS is destroyed. -- David Gibson | I'll have my music baroque, and my code david AT gibson.dropbear.id.au | minimalist, thank you. NOT _the_ _other_ | _way_ _around_! http://www.ozlabs.org/~dgibson
signature.asc
Description: Digital signature