Re: [PATCH V7 4/6] blk-mq: introduce .get_budget and .put_budget in blk_mq_ops

[Hannes Reinecke]

On 10/17/2017 03:29 AM, Ming Lei wrote:
> On Mon, Oct 16, 2017 at 01:30:09PM +0200, Hannes Reinecke wrote:
>> On 10/13/2017 07:29 PM, Ming Lei wrote:
>>> On Fri, Oct 13, 2017 at 05:08:52PM +0000, Bart Van Assche wrote:
>>>> On Sat, 2017-10-14 at 00:45 +0800, Ming Lei wrote:
>>>>> On Fri, Oct 13, 2017 at 04:31:04PM +0000, Bart Van Assche wrote:
>>>>>> On Sat, 2017-10-14 at 00:07 +0800, Ming Lei wrote:
>>>>>>> Actually it is in hot path, for example, lpfc and qla2xx's queue depth is 3,
>>>>>>
>>>>>> Sorry but I doubt whether that is correct. More in general, I don't know any modern
>>>>>> storage HBA for which the default queue depth is so low.
>>>>>
>>>>> You can grep:
>>>>>
>>>>> [ming@ming linux]$ git grep -n cmd_per_lun ./drivers/scsi/ | grep -E "qla2xxx|lpfc"
>>>>
>>>> Such a low queue depth will result in suboptimal performance for adapters
>>>> that communicate over a storage network. I think that's a bug and that both
>>>> adapters support much higher cmd_per_lun values.
>>>>
>>>> (+James Smart)
>>>>
>>>> James, can you explain us why commit 445cf4f4d2aa decreased LPFC_CMD_PER_LUN
>>>> from 30 to 3? Was that perhaps a workaround for a bug in a specific target
>>>> implementation?
>>>>
>>>> (+Himanshu Madhani)
>>>>
>>>> Himanshu, do you perhaps know whether it is safe to increase cmd_per_lun for
>>>> the qla2xxx initiator driver to the scsi_host->can_queue value?
>>>
>>> ->can_queue is size of the whole tag space shared by all LUNs, looks it isn't
>>> reasonable to increase cmd_per_lun to .can_queue.
>>>
>> '3' is just a starting point; later on it'll be adjusted via
>> scsi_change_depth().
>> Looks like it's not working correctly with blk-mq, though.
> 
> At default, in scsi_alloc_sdev(), q->queue_depth is set as
> host->cmd_per_lun. You are right, q->queue_depth can be adjusted
> later too.
> 
> q->queue_depth is respected in scsi_dev_queue_ready().
> .cmd_per_lun defines the max outstanding cmds for each lun, I
> guess it is respected by some hardware inside.
> 
No, this is purely a linux abstraction. Nothing to do with the hardware.

> For example, I remembered that on lpfc q->queue_depth is 30 because
> the default 'lpfc_lun_queue_depth' is 30. And its .cmd_per_lun is 3.
> Per my observation, this .cmd_per_lun limit is still workable.
> 
Again, these are just some pre-defined values to avoid I/O starvation
when having several LUNs. _if_ we can guarantee I/O fairness between
several (hundreds!) devices all sharing the same tagspace we wouldn't
need these variables.

Cheers,

Hannes
-- 
Dr. Hannes Reinecke		   Teamlead Storage & Networking
hare@xxxxxxx			               +49 911 74053 688
SUSE LINUX GmbH, Maxfeldstr. 5, 90409 NÃrnberg
GF: F. ImendÃrffer, J. Smithard, J. Guild, D. Upmanyu, G. Norton
HRB 21284 (AG NÃrnberg)