[dm-devel] Re: dm snapshot: stop merging using a completion

[Mike Snitzer]

On Sat, Dec 05 2009 at  9:01pm -0500,
Mikulas Patocka <mpatocka at redhat.com> wrote:

> On Sat, 5 Dec 2009, Mike Snitzer wrote:
> 
> > Switch stop_merge() from using a busy loop to a completion event.
> > 
> > stop_merge() now requests merging be shutdown using the
> > 'merge_completion' pointer (instead of the 'merge_shutdown' flag).  This
> > is accomplished by testing if 'merge_completion' is not NULL in
> > snapshot_merge_process().  stop_merge() allocates its completion on the
> > stack and assigns it to the 'merge_completion' pointer in the snapshot.
> > 'merge_completion' is protected by the snapshot's lock.
> > 
> > Also changed the 'merge_running' flag from int to atomic_t.
> 
> No, there's a bug:
> 
> >  static void stop_merge(struct dm_snapshot *s)
> >  {
> > -	while (s->merge_running) {
> > -		s->merge_shutdown = 1;
> > -		msleep(1);
> > +	DECLARE_COMPLETION_ONSTACK(merge_stopped);
> > +	if (atomic_read(&s->merge_running)) {
> 
> --- if the merge stops exactly at this point (because it gets finished or 
> because of an i/o error), we are waiting for a completion that will be 
> never signalled.

Yes, valid point.  But for this rare corner case we could just use
wait_for_completion_timeout() with a fairly large timeout; like 30 sec?
That actually isn't a great option (racey)...

How about if the 'shut:' code paths also checked for s->merge_completion
and complete() if it is not NULL?  Which means that check and related
complete() code would become a function.

> > +		down_write(&s->lock);
> > +		s->merge_completion = &merge_stopped;
> > +		up_write(&s->lock);
> > +		wait_for_completion(&merge_stopped);
> >  	}
> > -	s->merge_shutdown = 0;
> >  }
> >  
> >  /*
> 
> For Alasdair: do you get the problem? If I write it with msleep() 
> correctly, you keep on complaining how unclean it is --- if it is written 
> with completions and it is wrong (because they are just harder to use 
> correctly than simple variables and msleep), you tend to support it. Now 
> you see in practice how complex constructs tend to trigger bugs.
> 
> Mike: I thought that the completion would be in struct dm_snapshot. But 
> maybe, try it with wait_on_bit / wake_up_bit / test_bit / set_bit etc., it 
> may be easier than completions.

I can look at it; but I think using a completion can work.

Mike