Re: [RFC PATCH 4/5] RCU: Add TASK_RCU_OFFSET

[Lai Jiangshan <laijs@cn.fujitsu.com>]

On 03/30/2011 08:47 AM, Paul E. McKenney wrote:
> On Tue, Mar 29, 2011 at 03:01:19PM -0700, H. Peter Anvin wrote:
>> On 03/29/2011 02:47 PM, Paul E. McKenney wrote:
>>> On Tue, Mar 29, 2011 at 02:32:30PM -0700, H. Peter Anvin wrote:
>>>> On 03/29/2011 02:31 PM, Paul E. McKenney wrote:
>>>>>>
>>>>>> I have to say that if we have to use hardcoded offsets in C then we have
>>>>>> bigger problems.
>>>>>
>>>>> In this case, the offsets are mechanically generated from the structure
>>>>> definitions.
>>>>>
>>>>> Or am I missing your point?
>>>>
>>>> Yes.  The point is if we have to pull out these kinds of hacks in *C*
>>>> code, we are doing it wrong.  Not just a little wrong, but completely
>>>> and totally bonkers wrong.
>>>
>>> OK, maybe we are doing it wrong.
>>>
>>> But in that case, how do you suggest restructuring include/linux/sched.h
>>> so that struct task_struct can be safely included everywhere
>>> rcu_read_lock() and friends are invoked?  Or, on the other hand,
>>> what should we be doing so that we don't need to include task_struct
>>> everywhere?
>>
>> Lai's text doesn't give any hint as to the specific nature of the
>> conflict, which makes it hard to come up with a better alternative
>> without having to rediscover the problem from first principles.
>> However, a somewhat logical assumption is that the problem is that
>> struct task_struct includes struct rcu_head, in which case the easiest
>> thing to do is almost certainly to move the definition of struct
>> rcu_head to its own header file, <linux/rcuhead.h>, and include that in
>> <linux/sched.h>, which should make it possible to include
>> <linux/sched.h> in <linux/rcupdate.h>.
> 
> I believe that there are other circular dependencies -- there certainly
> were a few years back -- but I will defer to Lai.
> 
> 							Thanx, Paul
> 

Yes, there are other circular dependencies, <linux/sched.h> includes
many files which include or indirectly include <linux/rcupdate.h>
for struct rcu_head or RCU apis. There are too many to be split all.

Even we just create a <linux/task_struct_def.h>, it also needs to include
many files which have included or indirectly included <linux/rcupdate.h>
for struct rcu_head already. It is still not a easy work to split them,
it still requires many subsystem maintainers to participate in.

RCU is one of the most import subsystem,
I think it is worth to do such "wrong" way.

Thanks,
Lai

P.S.

Circular dependencies is one of the original sins of C language.
If it is a good change for lots of linux hackers participating in
to split all un-split kernel headers, I'd like to do so too.