On 04/02/2015 18:25, Emilio G. Cota wrote:
> On Wed, Feb 04, 2015 at 11:01:00 +0100, Paolo Bonzini wrote:
>> On 03/02/2015 23:08, Emilio G. Cota wrote:
>>> This matches the semantics of liburcu.
>>
>> This is not necessary. The two sets of macros are exactly the same, so
>> it's okay to use atomic_rcu_read/write.
>
> They're not exactly the same, otherwise the patch would be trivial.
You're right, I was imprecise---I meant they are interoperable. You can
use atomic_rcu_read/write together with liburcu, you do not need to use
the liburcu-provided rcu_dereference/rcu_assign_pointer.
Paolo
> The difference can be seen in the change to the macros' documentation:
>
> On Tue, Feb 03, 2015 at 17:08:18 -0500, Emilio G. Cota wrote:
>> This matches the semantics of liburcu.
>> --- a/docs/rcu.txt
>> +++ b/docs/rcu.txt
> (snip)
>> - typeof(*p) atomic_rcu_read(p);
>> + typeof(p) rcu_dereference(p);
> (snip)
>> - void atomic_rcu_set(p, typeof(*p) v);
>> + void rcu_assign_pointer(p, typeof(p) v);
>
> The liburcu macros take a variable (usually a pointer, hence the
> macros' names, but unsigned long is also common), not its address.
>
> These changes require modifications to the calling code as well, e.g.
> memory.c:
>
> struct AddressSpace {
> ...
> /* Accessed via RCU. */
> struct FlatView *current_map;
> ...
> };
> ...
> struct FlatView *view;
> ...
> - view = atomic_rcu_read(&as->current_map);
>
>
>
> + view = rcu_dereference(as->current_map);
> ...
> - atomic_rcu_set(&as->current_map, new_view);
>
>
>
> + rcu_assign_pointer(as->current_map, new_view);
>
> Thanks,
>
> Emilio
>
