On 15 March 2017 at 12:49, Zhong, Yang <yang.zh...@intel.com> wrote:
> So we can omit the cpu-memory address space allocated in the
> address_space_init_shareable(), which will same around 3M physical memory.
I see what you want to do...
> Maybe the below patch is much better, please help review, thanks!
>
> diff --git a/memory.c b/memory.c
> index 64b0a60..230f2cb 100644
> --- a/memory.c
> +++ b/memory.c
> @@ -2422,7 +2422,7 @@ AddressSpace *address_space_init_shareable(MemoryRegion
> *root, const char *name)
> AddressSpace *as;
>
> QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
> - if (root == as->root && as->malloced) {
> + if (root == as->root && (as->malloced || as ==
> &address_space_memory)) {
> as->ref_count++;
> return as;
> }
...but all the stuff I said below applies to this patch.
> It looks wrong to me. address_space_memory is not allocated via
> address_space_init_shareable(), so it's not correct to treat it that way
> (we're implicitly relying on it never being destroyed). It works by accident,
> not by design.
>
> If we want to support sharing of address spaces which are constant and exist
> for the lifetime of QEMU then we should probably do it with a new function
> something like
> address_space_init_static_shareable()
> which marks the AS as being (1) shareable and (2) invalid to ever try to
> destroy. Then you could use that for both address_space_memory and
> address_space_io.
We should not be special casing address_space_memory here:
if you want a mechanism for allowing static "exists for
lifetime of QEMU" address spaces to be included in the
set which can be shared, then you need to create one.
thanks
-- PMM
