Re: [Qemu-devel] [PATCH 02/16] exec.c: Allow target CPUs to define multiple AddressSpaces

["Edgar E. Iglesias" <edgar.iglesias@gmail.com>]

On Fri, Nov 06, 2015 at 01:34:44PM +0000, Peter Maydell wrote:
> On 6 November 2015 at 13:21, Edgar E. Iglesias <edgar.iglesias@gmail.com> wrote:
> > On Thu, Nov 05, 2015 at 06:15:44PM +0000, Peter Maydell wrote:
> >> Allow multiple calls to cpu_address_space_init(); each
> >> call adds an entry to the cpu->ases array at the specified
> >> index. It is up to the target-specific CPU code to actually use
> >> these extra address spaces.
> >>
> >> Since this multiple AddressSpace support won't work with
> >> KVM, add an assertion to avoid confusing failures.
> >>
> >> Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
> 
> >> +    if (asidx >= cpu->num_ases) {
> >> +        if (cpu->num_ases == 0) {
> >> +            cpu->cpu_ases = g_new(CPUAddressSpace, asidx + 1);
> >> +        } else {
> >> +            cpu->cpu_ases = g_renew(CPUAddressSpace, cpu->cpu_ases, asidx + 1);
> >
> > IIUC, g_renew may move the entire cpu_ases area. The internals of
> > memory_listener_register (called below) seem to put away the pointers to listeners
> > so a renew+move would leave invalid pointers to listeners in memory.c wouldn't it?
> >
> > There are various ways of solving this, (e.g dynamic allocation of the listener,
> > static allocation of the cpu_ases, invalidate all listeners and restore them after
> > each as init and more). I'm sure you'll figure something out.
> 
> Oops, yes, you're right.
> 
> Maybe we should just have the target CPU say in advance what the
> maximum number of AddressSpaces it will have is -- my expectation
> is that this will be (a) small (b) known in advance anyway.

Yes, that sounds good to me.

Cheers,
Edgar