On Wed, 29 Apr 2015 18:02:58 +0200 Peter Zijlstra <peterz@infradead.org>
wrote:

> On Wed, Apr 29, 2015 at 03:35:12PM +0200, Christoph Hellwig wrote:
> > On Wed, Apr 29, 2015 at 07:25:30AM +1000, NeilBrown wrote:
> > > As bdi_set_min_ratio doesn't touch bdi->dev, there seems to be no need for
> > > the test, or the warning.
> > > 
> > > I wonder if it would make sense to move the bdi_set_min_ratio() call to
> > > bdi_destroy, and discard bdi_unregister??
> > > There is a comment which suggests bdi_unregister might be of use later, but
> > > it might be best to have a clean slate in which to add whatever might be
> > > needed??
> > 
> > This seems fine to me from the block dev point of view.  I don't really
> > understand the bdi_min_ratio logic, but Peter might have a better idea.
> 
> Ah, that was a bit of digging, I've not looked at that in ages :-)
> 
> So if you look at bdi_dirty_limit()'s comment:
> 
>  * The bdi's share of dirty limit will be adapting to its throughput and
>  * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
> 
> So the min_ratio is a minimum guaranteed fraction of the total
> throughput.
> 
> Now the problem before commit ccb6108f5b0b ("mm/backing-dev.c: reset bdi
> min_ratio in bdi_unregister()") was that since bdi_set_min_ratio()
> keeps a global sum of bdi->min_ratio, you need to subtract from said
> global sum when taking the BDI away. Otherwise we loose/leak a fraction
> of the total throughput available (to the other BDIs).
> 
> Which is what that bdi_set_min_ratio(bdi, 0) in unregister does. It
> resets the min_ratio for the bdi being taken out and frees up the min
> allocated bandwidth for the others.
> 
> So I think moving that do destroy would be fine; assuming the delay
> between unregister and destroy is typically 'short'. Because without
> that you can 'leak' this min ratio for extended periods which means the
> bandwidth is unavailable for other BDIs.
> 
> Does that make sense?

Your assessment is almost exactly what I had come up with, so it definitely
makes sense :-)
'destroy' does come very shortly after 'unregister' (and immediately before
'blk_put_queue' which actually frees the struct).  However the driving force
for this patch was a desire to move blk_cleanup_queue(), which calls
'destroy', earlier.  So the net result is that bdi_set_min_ratio will be
called slightly sooner.

Thanks,
NeilBrown