Re: [PATCH v7 3/4] block: implement runtime pm strategy

[Aaron Lu <aaron.lu@intel.com>]

On Thu, Jan 17, 2013 at 10:11:31AM -0500, Alan Stern wrote:
> On Thu, 17 Jan 2013, Aaron Lu wrote:
> 
> > > > @@ -515,6 +516,27 @@ void elv_bio_merged(struct request_queue *q, struct request *rq,
> > > >  		e->type->ops.elevator_bio_merged_fn(q, rq, bio);
> > > >  }
> > > >  
> > > > +#ifdef CONFIG_PM_RUNTIME
> > > > +static void blk_pm_requeue_request(struct request *rq)
> > > > +{
> > > > +	if (!(rq->cmd_flags & REQ_PM))
> > > > +		rq->q->nr_pending--;
> > > > +}
> > > 
> > > You don't check q->dev here.  That's okay, but it means that
> > > q->nr_pending will be meaningless or wrong if any I/O takes place
> > > before blk_pm_runtime_init is called.
> > 
> > Right, so I had better check q->dev here too.
> > 
> > > 
> > > Therefore the kerneldoc for blk_pm_runtime_init should mention that it
> > > must not be called after any requests have been submitted.  Also
> > 
> > So with the q->dev check added above, I believe this is not needed.
> 
> No, it is still needed.  With the q->dev check added, q->nr_pending 
> will always be 0 before blk_pm_runtime_init is called.  But if I/O is 
> taking place then the number of pending requests really isn't 0.

Right, we can't allow I/O to happen when blk_pm_runtime_init is
executing, or we will have an incorrect nr_pending.

> 
> Either you have to make sure the q->nr_pending is always correct, even 
> when runtime PM isn't being used, or else the caller has to make sure 
> that no I/O takes place before blk_pm_runtime_init is called.

I think we can say:
blk_pm_runtime_init can't be called after any requests have been
submitted but not finished.
Sounds more accurate?

Thanks,
Aaron

> 
> > ------------
> > > mention that blk_pm_runtime_init enables runtime PM for q->dev, so the
> > > caller shouldn't do it.
> > 
> > I may misunderstandd this in last email.
> > 
> > We didn't enable runtime PM for the device in blk_pm_runtime_init, just
> > some auto suspend related settings.
> 
> Sorry, yes, you are right.  In fact, the caller has to make sure that 
> runtime PM is enabled before calling blk_pm_runtime_init; otherwise the 
> autosuspend timer might not work right.
> 
> Alan Stern
>