[Pkg-iscsi-maintainers] Bug#748940: iscsitarget: "service iscsitaget stop" fails when under load

[Craig Barnes]

On 22/05/14 17:45, Ritesh Raj Sarraf wrote:
> Please keep the bug email in the recipient list.
>
>
> On 05/22/2014 09:37 PM, Craig Barnes wrote:
>>> Could you please explain what is "heavy load" here? Is it iSCSI traffic ?
>>
>> To simulate heavy load ( it does not need to be too heavy ), I initiated
>> 4 dd operations to files on the device, then issued the stop.
>>
>
> I am assuming you are writing to the iscsi devices. If yes, then
> there's no reason for this to not fail.

Yes I am writing to the iscsi device, the devices get removed as
expected, the service initiates a stop, only the kernel module unload
fails, I assume waiting for the service to finish or disks to sync.

Is there any reason why this shouldn't successfully stop?

>
>>>
>>>>
>>>> # service iscsitarget stop
>>>> [ ok ] Removing iSCSI enterprise target devices: :.
>>>> [ ok ] Stopping iSCSI enterprise target service: :.
>>>> [....] Removing iSCSI enterprise target modules: :FATAL: Module
>>>> iscsi_trgt is in use.
>>>>
>>>> Changing etc/init/init.debian to use `rmmod --wait` instead of
>>>> `modprobe -r` corrects this issue.
>>>
>>> We use 'modprobe -r' to ensure that dependencies also get unloaded. With
>>> 'rmmod --wait', wouldn't it hang the shutdown sequence ?
>>
>> I wasn't able to determine any dependant libraries on my system, if
>> there are any in other packages that depend on iscsi_trgt then I would
>> be happy to test this and improve my solution.
>>
>
> Yes. THis one does not have any direct dependencies. As practice, I use
> modprobe.

Ok, so
   a) what is the worst thing that could happen using rmmod --wait
   b) is there a better way to do this? maybe wait 2 secs and try again
with modprobe -r?
   c) if we fail at the module shouldn't we bring everything else back
again?

>
>> Our deployment has many targets so this is not practical for our use case.
>>
>> I can see from the other init implementations that init.debian is the
>> only one that unloads the modules, is this still necessary?
>>
>
> The userspace daemon is very basic. The sessions are handled in the
> kernel. So when you ask iscsitarget to stop, you want the session also
> terminated, for which, we need to unload the module.
>
> I haven't confirmed it on the iscsitarget code but for open-iscsi (the
> initiator) that is the case.
>
> Yes. That is the case. Look at conn.c in the sources.

Sorry my c is a bit rusty and I am no kernel dev.  But I understand and
appreciate your explanation.

However I did find this...
http://sourceforge.net/p/iscsitarget/code/343/