Re: [PATCH 2/2] ufs: don't use the fair tag sharings

[Yu Kuai]

Hi,

在 2023/05/16 23:12, Bart Van Assche 写道:
> On 5/12/23 20:09, Yu Kuai wrote:
> > 在 2023/05/13 2:12, Bart Van Assche 写道:
> > > The fair tag sharing algorithm has a negative impact on all SCSI 
> > > devices with multiple logical units. This is because logical units 
> > > are considered active until (request timeout) seconds have elapsed 
> > > after the logical unit stopped being used (see also the 
> > > blk_mq_tag_idle() call in blk_mq_timeout_work()). UFS users are hit 
> > > by this because UFS 3.0 devices have a limited queue depth (32) and 
> > > because power management commands are submitted to a logical unit 
> > > (WLUN). Hence, it happens often that the block layer "active queue" 
> > > counter is equal to 2 while only one logical unit is being used 
> > > actively (a logical unit backed by NAND flash). The performance 
> > > difference between queue depths 16 and 32 for UFS devices is 
> > > significant.
> >
> > We meet similiar problem before, but I think remove tag fair sharing
> > might cause some problems, because get tag is not fair currently, for
> > example 2 devices share 32 tag, while device a issue large amount of
> > io concurrently, and device b only issue one io, in this case, if fair
> > tag sharing is removed, device b can get bad io latency.
> >
> > By the way, I tried to propose a way to workaround this by following:
> >
> > 1) disable fair tag sharing untill get tag found no tag is avaiable;
> > 2) enable fair tag sharing again if the disk donesn't faild to get tag
> > for a period of time;
> >
> > Can this approch be considered?
>
> I'm afraid that this approach won't help for the UFS driver since it is 
> likely that all tags are in use by a single logical unit during an IOPS 
> test. Hence, fair sharing would be enabled even when we don't want it to 
> be enabled.

It's right my original method is not flexible.

> I propose that we switch to one of these two approaches:

How about a smoothing method that the device with more io will share
more tag, and each device will get at least one tag?

Thanks,
Kuai

> * Either remove the fair tag sharing code entirely and rely on the 
> fairness mechanism provided by the sbitmap code. I'm referring to how 
> __sbitmap_queue_wake_up() uses the wake_index member variable.
> * Or make the behavior of the fairness algorithm configurable from user 
> space. One possible approach is to make the proportion of tags for a 
> logical unit / NVMe namespace configurable via sysfs. This will allow to 
> reduce the number of tags for the WLUN of UFS devices.
>
> Thanks,
>
> Bart.
>
>
> .