Re: [PATCH v4 6/6] scsi: core: Improve IOPS in case of host-wide tags

[Bart Van Assche <bart.vanassche@gmail.com>]

On 12/19/25 3:06 PM, Damien Le Moal wrote:
> On 12/20/25 02:35, Bart Van Assche wrote:
>> On 12/16/25 7:24 PM, Damien Le Moal wrote:
>>> On 12/17/25 07:30, Bart Van Assche wrote:
>>>> The SCSI core uses the budget map to restrict the number of commands
>>>> that are in flight per logical unit. That limit check can be left out if
>>>> host->cmd_per_lun >= host->can_queue and if the host tag set is shared
>>>> across all hardware queues or if there is only one hardware queue  Since
>>>
>>> Missing a period at the end of the sentence (before Since). But more
>>> importantly, this does not explain why the above is true, and frankly, I do not
>>> see it...
>> Hi Damien,
>>
>> The purpose of the SCSI device budget map is to prevent that the queue
>> depth limit for that SCSI device is exceeded. If there is only a single
>> hardware queue or there is a host-wide tag set and host->cmd_per_lun is
>> identical to host->can_queue, it is not possible that the queue depth
>> for a single SCSI device is exceeded and hence the SCSI device budget
>> map is not needed.
> 
> Still very confusing because as far as I understand things, a host is not
> necessarily a LUN/block device (you can have several devices/LUNs on the same
> host). So in general host->can_queue != device max queue depth. Also, I am not
> entirely sure if host->cmd_per_lun and max queue depth of the LUN are the same
> thing, given that SCSI does not define a maximum device queue depth...

Hi Damien,

There are important use cases, e.g. the UFS driver, where
host->can_queue is identical to the maximum queue depth per logical
unit. A single UFS device typically supports multiple logical units.

>> Please help with reviewing the ATA patch in this series.
> 
> For AHCI, we are dealing with single queue devices, always. For this case, I do
> not think that the scsi budget is needed since we will always have scsi tag ==
> ATA tag, between 0 and 31. So if you can allocate a tag, you can always submit
> the command.
> 
> But for libsas HBAs, I am not sure at all if what you did is solid/works,
> because I still do not understand it. Please provide more detailed explanations
> in your code (comments) and commit messages to better explain what you are doing
> is safe.

I plan to modify scsi_needs_budget_map() in patch 6/6 such that SCSI
hosts that define a .change_queue_depth method and/or that set
.track_queue_depth. This will prevent that this optimization applies to
libsas HBAs. From include/scsi/libsas.h:

#define __LIBSAS_SHT_BASE						\
	[ ... ]
	.change_queue_depth		= sas_change_queue_depth,	\
	[ ... ]

Thanks,

Bart.