On Thu, Sep 18, 2025 at 09:47:07PM +0900, Akihiko Odaki wrote: > On 2025/09/18 21:39, Cédric Le Goater wrote: > > Hello Akihiko, > > > > On 9/17/25 14:56, Akihiko Odaki wrote: > > > Based-on: <[email protected]> > > > ("[PATCH 00/14] Fix memory region use-after-finalization") > > > > > > Make AddressSpaces QOM objects to ensure that they are destroyed when > > > their owners are finalized and also to get a unique path for debugging > > > output. > > > > > > Suggested by BALATON Zoltan: > > > https://lore.kernel.org/qemu-devel/cd21698f-db77-eb75-6966- > > > [email protected]/ > > > > > > Signed-off-by: Akihiko Odaki <[email protected]> > > > > I wonder if this is going to fix an issue I was seeing a while ago > > in the FSI models. I couldn't find a clean way to avoid corrupting > > memory because of how the address_space was created and later on > > destroyed. See below, > > Partially, but this is insufficient. > > The first problem is that AddressSpace suffers from circular references the > following series solves: > https://lore.kernel.org/qemu-devel/[email protected]/ > "[PATCH v2 0/3] memory: Stop piggybacking on memory region owners" > > Another problem is that RCU is not properly waited. This is left to future > work.
Just to mention, Peter Maydell just posted a series for fixing AS destructions here: https://lore.kernel.org/qemu-devel/[email protected]/ IIUC it should also work for FSI, if FSI can convert to use dynamically allocated AddressSpaces (with/without QOMify; as Akihiko pointed out correctly, these should be orthogonal), then provide a proper unrealize() of the bus device to invoke address_space_destroy_free(). Thanks, -- Peter Xu
