Re: [Qemu-devel] qdev_get_machine() can return something non-NULL but not TYPE_MACHINE

[Igor Mammedov]

On Tue, 5 Sep 2017 10:08:01 +0100
Peter Maydell <address@hidden> wrote:

> On 4 September 2017 at 20:38, Igor Mammedov <address@hidden> wrote:
> > On Mon, 4 Sep 2017 17:36:59 +0100
> > Peter Maydell <address@hidden> wrote:
> >  
> >> I just got bitten by qdev_get_machine()'s behaviour on the user-only
> >> emulators, where it can return something that isn't NULL and isn't
> >> an instance of TYPE_MACHINE either.  
> > user-only shouldn't get to qdev_get_machine() at all,
> > issue probably in container_get().  
> 
> I need it in cpu_common_realizefn(), for
> http://patchwork.ozlabs.org/patch/797940/
Link might be broken (unable to connect to server)

Anyways I'd avoid using machine from cpu_*_realizefn(),
instead of I'd add property to CPU that has needed data
and set it from board code. Should work fine for *-user targets
and maintain clear separation of device impl. and board details. 

> 
> > I'd try to fix wrong user if possible and maybe add ifdef build failure
> > to qdev_get_machine() so it would not build in user mode.  
> 
> Can't ifdef, that source file is built once for all targets.
> 
> My fix (which I intend to send to the list today) is to make
> it do the object_dynamic_cast() check -- if that doesn't give
> a TYPE_MACHINE then we're in user mode and don't need to set
> ignore_memory_transaction_failures on the cpu object anyway.
> 
> >> It looks like maybe this can happen in some cases in softmmu too,
> >> judging by the way that qdev_get_hotplug_handler() does an
> >> object_dynamic_cast() check that it really got back a TYPE_MACHINE.  
> > As I recall only bus or machine provide hotplug_handler currently,
> > but it's possible to extend to other objects if we find use-case.
> >
> > We could do static cast to machine instead dynamic there but
> > in hotplug case it will abort QEMU if error happens,
> > hence dynamic check to avoid be more resilient during hotplug.
> > (well, if qdev_get_machine() returns not machine during startup
> > we would be screwed anyways, but that should break much earlier)  
> 
> If this can't ever happen then we should be aborting; that's
> the idea behind the cast macros doing assertions. I'm not
> sure hotplug needs to be special here if it doesn't have
> a genuine reason to think it might get back something of
> the wrong type.
Yep, we should abort.

> 
> thanks
> -- PMM