[Ocfs2-devel] avoid being purged when queued for assert_master

[Sunil Mushran]

The last email you said it reproduced. Now you say it did not.
I'm confused.

On 10/12/2011 07:13 PM, Wengang Wang wrote:
> On 11-10-12 19:11, Sunil Mushran wrote:
>> That's what ovm does. Have you reproduced it with ovm3 kernel?
>>
> No, I have no reproductions.
>
> thanks,
> wengang.
>> On 10/12/2011 07:07 PM, Wengang Wang wrote:
>>> On 11-10-13 09:51, Wengang Wang wrote:
>>>> On 11-10-12 18:47, Sunil Mushran wrote:
>>>>> I meant master_request (not query). We set refmap _before_
>>>>> asserting. So that should not happen.
>>>> Why can't the remote node requested deref (DLM_DEREF_LOCKRES_MSG)?
>>> The problem can easily happen on this dlmfs useage:
>>>
>>> reopen:
>>> 	open(create) /dlm/dirxx/filexx
>>> 	close	     /dlm/dirxx/filexx
>>> 	sleep 60
>>> 	goto reopen
>>>
>>>> thanks,
>>>> wengang.
>>>>> On 10/12/2011 06:02 PM, Wengang Wang wrote:
>>>>>> Hi Sunil,
>>>>>>
>>>>>> On 11-10-12 17:32, Sunil Mushran wrote:
>>>>>>> So you are saying a lockres can get purged before the node is asserting
>>>>>>> master to other nodes?
>>>>>>>
>>>>>>> The main place where we dispatch assert is during master_query.
>>>>>>> There we set refmap before dispatching. Meaning refmap will protect
>>>>>>> us from purging.
>>>>>>>
>>>>>>> But I think it could happen in master_requery, which only comes into
>>>>>>> play if a node dies during migration.
>>>>>>>
>>>>>>> Is that the case here?
>>>>>> I think this can mainly include the response for a master_request.
>>>>>> in dlm_master_request_handler(), the master node quques assert_master.
>>>>>> The node which requested a master_request knows the master by receving
>>>>>> response values. It doesn't need to wait until the assert_master come.
>>>>>> As you know, the asserting master work is done in a workqueue. And the
>>>>>> work item in it can be heavily delayed. So in the duriation from the
>>>>>> (old) master responding with "Yes, I am master" to it sending assert_master,
>>>>>> Anything can heppan, the worse case is the lockres on the (old) master
>>>>>> get purged and is remasted by another node. So in this case,
>>>>>> apparently, the old master shouldn't send the assert_master any longer.
>>>>>> To prevent that from happening, we should keep the lockres un-purged as
>>>>>> long as it's queued for master_request.
>>>>>>
>>>>>> #the problem is what my flush_workqueue patch tries to fix.
>>>>>>
>>>>>> thanks,
>>>>>> wengang.
>>>>>>
>>>>>>> On 10/12/2011 12:04 AM, Wengang Wang wrote:
>>>>>>>> Hi Sunil/Joel/Mark and anyone who has interest,
>>>>>>>>
>>>>>>>> This is not a patch but a discuss.
>>>>>>>>
>>>>>>>> Currently we have a problem:
>>>>>>>> When a lockres is still queued(in dlm->work_list) for sending an
>>>>>>>> assert_master(or in processing of sending), the lockres can't be
>>>>>>>> purged(removed from hash). there is no flag/state,on lockres its self,dinotes
>>>>>>>> this situation.
>>>>>>>>
>>>>>>>> The badness is that if the lockres is purged(surely not the owner at the
>>>>>>>> moment), and the assert_master is after the purge. it can confuse other
>>>>>>>> nodes. On another node, the owner now can be any other nodes, thus on
>>>>>>>> receiving the assert_master, it can trigger a BUG() because 'owner'
>>>>>>>> doesn't match.
>>>>>>>>
>>>>>>>> So we'd better to prevent the lockres from be purged when it's queued
>>>>>>>> for something(assert_master).
>>>>>>>>
>>>>>>>> Srini and I discussed some possible fixes:
>>>>>>>> 1) adding a flag to lockres->state.
>>>>>>>>     this does not work. A lockres can have multiple instances in the queue list.
>>>>>>>>     A simple flag is not safe. And the instances are not nested, so even
>>>>>>>>     saving a previous flags doesn't work. Neither can we merge the instances
>>>>>>>>     because they can be for different purposes.
>>>>>>>>
>>>>>>>> 2) checking if the lockres if queued before purging it.
>>>>>>>>    this works, but doesn't sounds good. it needs changes of current behaviour
>>>>>>>>    on the queue list.   Also, we have no idea on the performance of the checking
>>>>>>>>    (searching list).
>>>>>>>>
>>>>>>>> 3) making use of lockres->inflight_locks.
>>>>>>>>    this works, but seems to be a mis-use of inflight_locks.
>>>>>>>>
>>>>>>>> 4) adding a new member to lockres counting the queued time.
>>>>>>>>     this works and simple. but needs extra memory.
>>>>>>>>
>>>>>>>> I prefer to the 4).
>>>>>>>>
>>>>>>>> What's your idea?
>>>>>>>>
>>>>>>>> thanks,
>>>>>>>> wengang.
>>>>>>>>
>>>>>>>> _______________________________________________
>>>>>>>> Ocfs2-devel mailing list
>>>>>>>> Ocfs2-devel at oss.oracle.com
>>>>>>>> http://oss.oracle.com/mailman/listinfo/ocfs2-devel