SMP: ISR disable/enable vs. mutual exclustion

[Chris Johns]

Sebastian Huber wrote:
> On 2013-08-27 01:26, Chris Johns wrote:
>> Sebastian Huber wrote:
>>> On 2013-08-24 04:10, Chris Johns wrote:
>>>>> Thus the normal extract operation is not available on SMP. An extract
>>>>> variant which needs also the chain control as a parameter must be
>>>>> used.
>>>>
>>>> I think a node may need a back pointer to the chain control that
>>>> contains a
>>>> lock. I suspect we cannot have a single score chain control structure
>>>> for both
>>>> protected and unprotected operations and support the current extract.
>>>> I have
>>>> not looked at all the uses of extract in the code so I do not know if
>>>> the chain
>>>> control is available and even it is I think the node should handle
>>>> this.
>>>
>>> In order to use the back pointer you have to lock the chain, so this
>>> cannot be used. You have to know on which chain the node is.
>>>
>>
>> Yes you are correct. Should the locking be something the user of the
>> chains
>> should manage ? The chains is starting to become more and more complex.
>
> On single processor configuration the locking is done by the chains (ISR
> disable/enable) so I think there is no way out.

The RTEMS chains has protected, unprotected and now explicit and I am 
starting to wonder if the API has become too complicated for users with 
too many choices. Some choices work with SMP other do not. We really 
should have one version that works in all cases however this means 
breaking the existing API. On the other hand anyone not using the 
explicit API will break when moving code to an SMP target and only finds 
out once they move or try. They may have made design choices without 
being aware of the differences and possible issues. The one related to 
this change is the need to have the chain control available when 
extracting a node. This is a big difference.

I understand the migration and the need not to break an API however I am 
asking the question about the complexity of the API and if we should 
review what it is doing. With this change where you need the control 
chain to extract the node you have some data following the node and that 
could contain the lock and I suspect in a number of cases will contain a 
lock. I also wonder if the user of the chain can manage the locking 
better in some cases than letting the chain do it. The locking was fast 
before and now it is not as fast changing the dynamic.

Extending the API makes sense now and from a kernel change point of view 
however I am not convinced of the long term benefit when all the current 
rapid change settles and is forgotten.

> Please have a look at the patch here:
>
> https://www.rtems.org/pipermail/rtems-devel/2013-August/004370.html
>
> I had to add rtems_chain_explicit_extract() and
> rtems_chain_explicit_insert() routines.
>

I understand the reason you have added the extra call. We either need to 
break the existing API or extend the current one. Having two APIs is not 
helpful to users and we need just one. If there is an agree path to one 
API I would be happier. It could be a warning for non-SMP that the API 
will change.

Chris