Re: [PATCH 2/5] exec.c: use refcnt to protect device during dispatching

[Paolo Bonzini <pbonzini@redhat.com>]

Il 25/07/2012 10:12, liu ping fan ha scritto:
>>> >> +        qemu_rwlock_rdlock_devtree();
>>> >>          section = phys_page_find(page >> TARGET_PAGE_BITS);
>>> >> +        if (!(memory_region_is_ram(section->mr) ||
>>> >> +            memory_region_is_romd(section->mr)) && !is_write) {
>>> >> +            bk = get_backend(section->mr, addr);
>>> >> +            object_ref(bk);
>>> >> +        } else if (!memory_region_is_ram(section->mr) && is_write) {
>>> >> +            bk = get_backend(section->mr, addr);
>>> >> +            object_ref(bk);
>>> >> +        }
>>> >> +        qemu_rwlock_unlock_devtree();
>>> >>
>>> >>          if (is_write) {
>>> >>              if (!memory_region_is_ram(section->mr)) {
>>> >> @@ -3426,6 +3462,7 @@ void cpu_physical_memory_rw(target_phys_addr_t addr, uint8_t *buf,
>>> >>                      io_mem_write(section->mr, addr1, val, 1);
>>> >>                      l = 1;
>>> >>                  }
>>> >> +                object_unref(bk);
>> >
>> > Currently object_ref()/object_unref() are not atomic.  Will you send
> We obey the rule:
>            rdlock->search->ref_get,
>            wrlock->remove ->ref_put
> So can  it causes problem if object_ref()/object_unref()  are not atomic?

Yes, two CPUs can perform object_ref at the same time.

You can find a header file for atomic operations here:
https://github.com/bonzini/qemu/commit/atomics.patch

Paolo