[AGENT++] Agent++ Thread Safety

[Frank Fock]

Hi Megan,

The first approach prevents a thread B to get
a reference of an object locked by thread A
and then deleted by thread A. When thread A
would then release the lock, thread B would
use an invalid reference.

In the second scenario, the content of an
object cannot be modified concurrently.

Best regards,
Frank

Megan Smith wrote:
> Hi Frank,
> 
> In the second scenario, however, it is then possible for two different
> tables to be modified concurrently.  If this is so, then isn't it possible
> that the static variable _target_item in avl_map.cpp for example is being
> used by two different tables concurrently without any locking?
> 
> Thank you,
> Megan
> 
> 
> On 5/21/08, Frank Fock <fock at agentpp.com> wrote:
>> Hi Megan,
>>
>> Both version are correct depending on the
>> usage scenario:
>>
>> If an existing object (i.e., row,table,leaf) is
>> about to be deleted, you need:
>>
>>  mib->lock_mib();
>>  table->start_synch();
>>  ..some code modifying table content..
>>  table->end_synch();
>>  mib->unlock();
>>
>> If you will only modify an object's content or
>> add a row, for example, then
>>
>>  mib->lock_mib();
>>  table->start_synch();
>>  mib->unlock();
>>  ..some code modifying table content..
>>  table->end_synch();
>>
>> is sufficient.
>>
>> Best regards,
>> Frank
>>
>> Megan Smith wrote:
>>
>>> Hi Frank,
>>>
>>> Thank you for your quick response.
>>>
>>> In the documentation for mib->lock_mib() the mib lock is only held until
>>> the
>>> mib object lock is acquired:
>>>
>>> "While processing a (sub-)request the locking scheme is as follows: 1.
>>> This
>>> MIB lock is acquired. 2. The MIB object is identified, which is
>>> responsible
>>> for processing the sub-request. 3. The MIB objects lock (
>>> MibEntry::start_synch()<
>>> http://www.agentpp.com/doc3.x/classThreadManager.html#42bcd26c71747a3d01afd2a407b7f12e>)
>>>
>>> is acquired. 4. The MIB lock is released. 5. The sub-request is propagated
>>> to the MIB object. 6. When finished, the MIB objects lock is released."
>>>
>>> But the procedure you just suggested:
>>>
>>> mib->lock_mib();
>>> table->start_synch();
>>> ..some code modifying table content..
>>> table->end_synch();
>>> mib->unlock();
>>>
>>> doesn't unlock the mib itself until the last step.  Which one is correct?
>>> Or are they for different situations?
>>>
>>>
>>> Thanks!
>>> Megan
>>>
>>>
>>> On 5/20/08, Frank Fock <fock at agentpp.com> wrote:
>>>
>>>> Hi Megan,
>>>>
>>>> You are probably on the wrong track here. The AVL
>>>> trees should not be modified in the AGENT++ code
>>>> concurrently anywhere.
>>>>
>>>> Such modifications have to be protected by
>>>>
>>>> mib->lock_mib();
>>>> ..some code modifying mib registration (AVL tree)..
>>>> mib->unlock_mib();
>>>>
>>>> or
>>>>
>>>> mib->lock_mib();
>>>> table->start_synch();
>>>> ..some code modifying table content..
>>>> table->end_synch();
>>>> mib->unlock();
>>>>
>>>> Best regards,
>>>> Frank
>>>>
>>>> Megan Smith wrote:
>>>>
>>>> Hi,
>>>>> I posted this a few weeks ago and haven't heard any response yet.  We
>>>>> are
>>>>> using agent++ and have been experiencing segmentation faults from some
>>>>> corrupted memory in the mib tables during highly multi-threaded runs.
>>>>>  After
>>>>> running debugging utilities it appears that there are some static
>>>>> variables
>>>>> in the avl_map.cpp file that I can't find proper locks for anywhere.  Am
>>>>> I
>>>>> missing something?
>>>>>
>>>>> Thank you for your help,
>>>>> Megan
>>>>> _______________________________________________
>>>>> AGENTPP mailing list
>>>>> AGENTPP at agentpp.org
>>>>> http://lists.agentpp.org/mailman/listinfo/agentpp
>>>>>
>>>>> --
>>>> AGENT++
>>>> http://www.agentpp.com
>>>> http://www.mibexplorer.com
>>>> http://www.mibdesigner.com
>>>>
>>>> _______________________________________________
>>> AGENTPP mailing list
>>> AGENTPP at agentpp.org
>>> http://lists.agentpp.org/mailman/listinfo/agentpp
>>>
>> --
>> AGENT++
>> http://www.agentpp.com
>> http://www.mibexplorer.com
>> http://www.mibdesigner.com
>>
> _______________________________________________
> AGENTPP mailing list
> AGENTPP at agentpp.org
> http://lists.agentpp.org/mailman/listinfo/agentpp

-- 
AGENT++
http://www.agentpp.com
http://www.mibexplorer.com
http://www.mibdesigner.com