Re: [PATCH 04/16] scsi_dh_alua: Make stpg synchronous

[Mike Christie <michaelc@cs.wisc.edu>]

On 2/12/14 10:11 AM, Mike Christie wrote:
> On 2/12/14 9:29 AM, Hannes Reinecke wrote:
>> On 02/07/2014 02:54 AM, Mike Christie wrote:
>>> On 02/06/2014 07:24 PM, Mike Christie wrote:
>>>> On 01/31/2014 03:29 AM, Hannes Reinecke wrote:
>>>>> We should be issuing STPG synchronously as we need to
>>>>> evaluate the return code on failure.
>>>>>
>>>>> Signed-off-by: Hannes Reinecke <hare@suse.de>
>>>>
>>>> I think we need to also make dm-mpath.c use a non-ordered workqueue for
>>>> kmpath_handlerd. With this patch and the current ordered workqueue in
>>>> dm-mpath I think we will only be able to do one STPG at a time. I think
>>>> if we use a normal old non-ordered workqueue then we would be
>>>> limited to
>>>> have max_active STPGs executing.
>>>
>>> I goofed and commented in the order I saw the patches :) I take this
>>> comment back for this patch, because I see in 16/16 you added a new
>>> workqueue to scsi_dh_alua to do rtpgs and stpgs.
>>>
>>> For 16/16 though, do we want to make kmpath_aluad a non single threaded
>>> workqueue? It looks like max_active for single threaded is only one work
>>> at a time too.
>>>
>> Well, that was by intention.
>>
>> The workqueue will be triggered very infrequently (basically for
>> every path switch).
>> For implicit ALUA we just need to issue a RTPG to get the new path
>> status; there we might be suffering from single threaded behaviour.
>> But we need to issue it only once and it should be processed
>> reasonably fast.
>> For explicit ALUA we'll have to send an STPG, which has potentially
>> system-wide implications. So sending several to (supposedly)
>> different targets might actually be contraproductive, as the first
>> might have already set the status for the second call.
>> Here we most definitely _want_ serialisation to avoid superfluous STPGs.
>>
>
> What target is this?
>
> For our target it adds a regression. It only affects devices on the same
> port group. We then have multiple groups. Before the patch, we could
> failover/failback multiple devices in parallel. To do 64 devices it took
> about 3 seconds. With your patch it takes around 3 minutes.
>

Also, with your pg change patch, I think we can serialize based on group 
and it will do what you want and also allow us to do STPGs to different 
groups in parallel.