Re: [RFC, PATCH] kernel/rcu: add kfree_rcu

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Sat, Jan 03, 2009 at 03:59:40PM +0100, Manfred Spraul wrote:
> Paul E. McKenney wrote:
>> I would suggest instead using the bottom bit to differentiate between
>> these two cases, especially given that your approach makes it impossible
>> for callback processing to notice a NULL function pointer.  In addition,
>> this approach would allow different types of allocators to be specified
>> should this later prove to be helpful.  You should not have to shift the
>> offset because the rcu_head offset should always be a multiple of four
>> (or eight on 64-bit architectures).
>>   
> We must be careful: rcu_head might be always aligned, but are function 
> pointers always aligned?
> The x86 hardware allows arbitrary function pointers, I'm not sure what gcc 
> would do if '--falign-functions=0' is used.
> Are there other codepaths that assume that the lowest bit of a function 
> pointer is never set?

Good point.  I guess that we will have to worry about expandability when
and if the need arises.  I see a couple of ways of doing it, but they
are pretty ugly...

>> And we really are running into bugs that are detected by RCU's seeing a
>> null function pointer in the rcu_head structure at callback-invocation
>> time.  So, whatever encoding you choose, please leave a function-pointer
>> value of zero as an invalid value!
>>   
> Ok.
>
>>> --- a/kernel/rcutree.c
>>> +++ b/kernel/rcutree.c
>>> @@ -901,7 +901,7 @@ static void rcu_do_batch(struct rcu_data *rdp)
>>>  	while (list) {
>>>  		next = list->next;
>>>  		prefetch(next);
>>> -		list->func(list);
>>> +		rcu_docallback(list);
>>>     
>>
>> Good, you got all three of them!  ;-)
>>
>>   
> The patch was tested against rcutree ;-)

;-)

							Thanx, Paul