On 3/15/12 3:45 PM, Vladislav Bogdanov wrote:
> 15.03.2012 18:43, William Seligman wrote:
>> On 3/15/12 3:43 AM, Vladislav Bogdanov wrote:
>>> 14.03.2012 00:42, William Seligman wrote:
>>> [snip]
>>>> These were the log messages, which show that stonith_admin did its job and 
>>>> CMAN
>>>> was notified of the fencing: <http://pastebin.com/jaH820Bv>.
>>>
>>> Could you please look at the output of 'dlm_tool ls' and 'dlm_tool dump'?
>>>
>>> You probably have 'kern_stop' and 'fencing' flags there. That means that
>>> dlm is unaware that node is fenced.
>>
>> Here's 'dlm_tool ls' with both nodes running cman+clvmd+gfs2:
>> <http://pastebin.com/QrZtm1Ue>
>>
>> 'dlm_tool dump': <http://pastebin.com/UKWxx9Y4>
>>
>> For comparison, I crashed one node and looked at the same output on the
>> remaining node:
>> dlm_tool ls: <http://pastebin.com/cKVAGxsd>
>> dlm_tool dump: <http://pastebin.com/c0h0p22Q> (the post-crash lines begin at
>> 1331824940)
> 
> Everything is fine there, dlm correctly understands that node is fenced
> and returns to a normal state.
> 
> The only minor issue I see is that fencing took much time - 21 sec.

Hmm. My fencing agent works by toggling the power on a UPS. If all the agent
does is "action=off", it will cut power immediately. But if you tell it
"action=reboot", it will cut the load, wait 10 seconds, then turn the load back
on again; I found I needed that delay because otherwise the UPS might
confuse/overlap/ignore sequential commands.

Could this be an issue? I've noticed that my fencing agent always seems to be
called with "action=reboot" when a node is fenced. Why is it using 'reboot' and
not 'off'? Is this the standard, or am I missing a definition somewhere?

>>
>> I don't see the "kern_stop" or "fencing" flags. There's another thing I don't
>> see: at the top of 'dlm_tool dump' it displays most of the contents of my
>> cluster.conf file, except for the fencing sections. Here's my cluster.conf 
>> for
>> comparison: <http://pastebin.com/w5XNYyAX>
> 
> It also looks correct (I mean fence_pcmk), but I can be wrong here, I do
> not use cman.
> 
>>
>> cman doesn't see anything wrong in my cluster.conf file:
>>
>> # ccs_config_validate
>> Configuration validates
>>
>> But could there be something that's causing the fencing sections to be 
>> ignored?
>>
>>>>
>>>> Unfortunately, I still got the gfs2 freeze, so this is not the complete 
>>>> story.
>>>
>>> Both clvmd and gfs2 use dlm. If dlm layer thinks fencing is not
>>> completed, both of them freeze.
>>
>> I did 'grep -E "(dlm|clvm|fenc)" /var/log/messages' and looked at the time I
>> crashed the node: <http://pastebin.com/dvBtdLUs>. I see lines that indicate 
>> that
>> pacemaker and drbd are fencing the node, but nothing from dlm or clvmd. Does
>> this indicate what you suggest: Could dlm somehow be ignoring or overlooking 
>> the
>> fencing I put in? Is there any other way to check this?
> 
> No, dlm_controld (and friends) mostly uses different logging method -
> that is what you see in dlm_tool dump.

-- 
Bill Seligman             | Phone: (914) 591-2823
Nevis Labs, Columbia Univ | mailto://[email protected]
PO Box 137                |
Irvington NY 10533 USA    | http://www.nevis.columbia.edu/~seligman/

Attachment: smime.p7s
Description: S/MIME Cryptographic Signature

_______________________________________________
Linux-HA mailing list
[email protected]
http://lists.linux-ha.org/mailman/listinfo/linux-ha
See also: http://linux-ha.org/ReportingProblems

Reply via email to