Re: Crashing / unable to start domUs due to high number of luns?

[Nathan March <nathan@gt.net>]

On 1/31/2012 5:30 PM, Konrad Rzeszutek Wilk wrote:
> On Tue, Jan 31, 2012 at 01:42:23PM -0800, Nathan March wrote:
>> Hi All,
>>
>> We've got a xen setup based around a dell iscsi device with each xen
>> host having 2 lun's, we then run multipath on top of that. After adding
>> a couple new virtual disks the other day, a couple of our online stable
>> VM's suddenly hard locked up. Attaching to the console gave me nothing,
>> looked like they lost their disk devices.
>>
>> Attempting to restart them on the same dom0 failed with hot plug errors,
>> as did attempting to start them on a few different dom0's. After doing a
>> "multipath -F" to remove unused devices and manually bringing in just
>> the selected LUN's via "multipath diskname", I was able to successfully
>> start them. This initially made me think perhaps I was hitting some sort
>> of udev / multipath / iscsi device lun limit (136 luns, 8 paths per lun
>> = 1088 iscsi connections). Just to be clear, the problem occurred on
>> multiple dom0's at the same time so it definitely seems iscsi related.
>>
>> Now, a day later, I'm debugging this further and I'm again unable to
>> start VM's, even with all extra multipath devices removed. I rebooted
>> one of the dom0's and was able to successfully migrate our production
>> VM's off a broken server, so I've now got an empty dom0 that's unable to
>> start any vm's.
>>
>> Starting a VM results in the following in xend.log:
>>
>> [2012-01-31 13:06:16 12353] DEBUG (DevController:144) Waiting for 0.
>> [2012-01-31 13:06:16 12353] DEBUG (DevController:628)
>> hotplugStatusCallback /local/domain/0/backend/vif/35/0/hotplug-status.
>> [2012-01-31 13:07:56 12353] ERROR (SrvBase:88) Request wait_for_devices
>> failed.
>> Traceback (most recent call last):
>>    File "/usr/lib64/python2.6/site-packages/xen/web/SrvBase.py", line
>> 85, in perform
>>      return op_method(op, req)
>>    File
>> "/usr/lib64/python2.6/site-packages/xen/xend/server/SrvDomain.py", line
>> 85, in op_wait_for_devices
>>      return self.dom.waitForDevices()
>>    File "/usr/lib64/python2.6/site-packages/xen/xend/XendDomainInfo.py",
>> line 1237, in waitForDevices
>>      self.getDeviceController(devclass).waitForDevices()
>>    File
>> "/usr/lib64/python2.6/site-packages/xen/xend/server/DevController.py",
>> line 140, in waitForDevices
>>      return map(self.waitForDevice, self.deviceIDs())
>>    File
>> "/usr/lib64/python2.6/site-packages/xen/xend/server/DevController.py",
>> line 155, in waitForDevice
>>      (devid, self.deviceClass))
>> VmError: Device 0 (vif) could not be connected. Hotplug scripts not working.
>
> Was there anything in the kernel (dmesg) about vifs? What does your
> /proc/interrupts look like? Can you provide the dmesg that you get
> during startup. I am mainly looking for:
>
> NR_IRQS:16640 nr_irqs:1536 16
>
> How many guests are your running when this happens?
>
> One theory is that your are running out dom0 interrupts. Thought
> I *think* that was made dynamic in 3.0..
>
>
> Thought that does explain your iSCSI network wonky in the guest -
> was there anything in the dmesg when the guest started going bad?

Was running approximately 15 guests, although this persisted after 
migrating them off.

Nothing in dmesg (dom0 dmesg or xm dmesg) that looked abnormal at all, 
no references to vifs. Asides from the inability to start a VM, I 
couldn't seem to find any sort of error anywhere.

All the hosts show the same irq counts:[   34.903763] NR_IRQS:4352 
nr_irqs:4352 16

Unfortunately I'm not able to reproduce this now, but I've posted 
several different copies of /proc/interrupts here: 
http://pastebin.com/n7PWNeaZ

Full xm / kernel dmesg is uploaded here: http://pastebin.com/AtCvFBDS

>> [2012-01-31 13:07:56 12353] DEBUG (XendDomainInfo:3071)
>> XendDomainInfo.destroy: domid=35
>> [2012-01-31 13:07:58 12353] DEBUG (XendDomainInfo:2401) Destroying
>> device model
>>
>> I tried turning up udev's log level but that didn't reveal anything.
>> Reading the xenstore for the vif doesn't show anything unusual either:
>>
>> ukxen1 ~ # xenstore-ls /local/domain/0/backend/vif/35
>> 0 = ""
>>   bridge = "vlan91"
>>   domain = "nathanxenuk1"
>>   handle = "0"
>>   uuid = "2128d0b7-d50f-c2ad-4243-8a42bb598b81"
>>   script = "/etc/xen/scripts/vif-bridge"
>>   state = "1"
>>   frontend = "/local/domain/35/device/vif/0"
>>   mac = "00:16:3d:03:00:44"
>>   online = "1"
>>   frontend-id = "35"
>>
>> The bridge device (vlan91) exists, trying a different bridge doesn't
>> matter. Removing the VIF completely results in the same error for the
>> VBD. Adding debugging to the hotplug/network scripts didn't reveal
>> anything, it looks like they aren't even being executed yet. Nothing is
>> logged to xen-hotplug.log.
> OK, so that would imply the kernel hasn't been able to do the right
> thing. Hmm.
>
> What do you see when this happens with udev --monitor --kernel --udev
> --property ?

The remaining server I thought was doing this is apparently not (I was 
probably mistaken), so the 2 that were definitely doing it have been 
rebooted and I can't reproduce this at the moment.

I've been abusing a free server all morning with a loop to 
spawn/shutdown a VM repeatedly and flush / rescan multipath to see if I 
can reproduce this again. No luck so far unfortunately, but I'll keep 
trying.


>> The only thing I can think of that this may be related to, is gentoo
>> defaulted to a 10mb /dev which we filled up a few months back. We upped
>> the size to 50mb in the mount options and everything's been completely
>> stable since (~33 days). None of the /dev on the dom0's is higher than
>> 25% usage. Asides from adding the new luns, no changes have been made in
>> the past month.
>>
>> To try and test if removing some devices would solve anything, I tried
>> doing an "iscsiadm -m node --logout" and it promptly hard locked the
>> entire box. After a reboot, I was unable to reproduce the problem on
>> that particular dom0.
>>
>> I've still got 1 dom0 that's exhibiting the problem, if anyone is able
>> to suggest any further debugging steps?
>>
>> - Nathan
>> _______________________________________________
>> Xen-devel mailing list
>> Xen-devel@lists.xensource.com
>> http://lists.xensource.com/xen-devel


-- 
Nathan March<nathan@gt.net>
Gossamer Threads Inc. http://www.gossamer-threads.com/
Tel: (604) 687-5804 Fax: (604) 687-5806