Re: [PATCH 05/14] dm: set the static flush bio device on demand

[Dennis Zhou]

On Tue, Dec 04, 2018 at 03:28:07PM -0500, Mike Snitzer wrote:
> On Tue, Dec 04 2018 at  1:35pm -0500,
> Dennis Zhou <dennis@xxxxxxxxxx> wrote:
> 
> > The next patch changes the macro bio_set_dev() to associate a bio with a
> > blkg based on the device set. However, dm creates a static bio to be
> > used as the basis for cloning empty flush bios on creation. The
> > bio_set_dev() call in alloc_dev() will cause problems with the next
> > patch adding association to bio_set_dev() because the call is before the
> > bdev is associated with a gendisk (bd_disk is %NULL). To get around
> > this, set the device on the static bio every time and use that to clone
> > to the other bios.
> > 
> > Signed-off-by: Dennis Zhou <dennis@xxxxxxxxxx>
> > Cc: Alasdair Kergon <agk@xxxxxxxxxx>
> > Cc: Mike Snitzer <snitzer@xxxxxxxxxx>
> > ---
> >  block/bio.c     |  1 +
> >  drivers/md/dm.c | 12 +++++++++++-
> >  2 files changed, 12 insertions(+), 1 deletion(-)
> > 
> > diff --git a/block/bio.c b/block/bio.c
> > index 452b8e79b998..41ebb3f8e2fc 100644
> > --- a/block/bio.c
> > +++ b/block/bio.c
> > @@ -2021,6 +2021,7 @@ void bio_disassociate_blkg(struct bio *bio)
> >  		bio->bi_blkg = NULL;
> >  	}
> >  }
> > +EXPORT_SYMBOL_GPL(bio_disassociate_blkg);
> >  
> >  /**
> >   * __bio_associate_blkg - associate a bio with the a blkg
> > diff --git a/drivers/md/dm.c b/drivers/md/dm.c
> > index a733e4c920af..a2d6f8b33d23 100644
> > --- a/drivers/md/dm.c
> > +++ b/drivers/md/dm.c
> > @@ -1417,10 +1417,21 @@ static int __send_empty_flush(struct clone_info *ci)
> >  	unsigned target_nr = 0;
> >  	struct dm_target *ti;
> >  
> > +	/*
> > +	 * Empty flush uses a statically initialized bio as the base for
> > +	 * cloning, &md->flush_bio.  However, blkg association requires that
> 
> Would prefer:
> "Empty flush uses a statically initialized bio, &md->flush_bio, as the
> base for cloning.  ..."
> 

Cool, that reads better.

> > +	 * a bdev is associated with a gendisk, which doesn't happen until the
> > +	 * bdev is opened.  So, blkg association is done at issue time of the
> > +	 * flush rather than when the device is created in dm_alloc().
> 
> Another nit but I think you mean "alloc_dev()" here .......^
> 

Yeah, I did mean alloc_dev(). Fixed.

> > +	 */
> > +	bio_set_dev(ci->bio, ci->io->md->bdev);
> > +
> >  	BUG_ON(bio_has_data(ci->bio));
> >  	while ((ti = dm_table_get_target(ci->map, target_nr++)))
> >  		__send_duplicate_bios(ci, ti, ti->num_flush_bios, NULL);
> >  
> > +	bio_disassociate_blkg(ci->bio);
> > +
> >  	return 0;
> >  }
> >  
> > @@ -1939,7 +1950,6 @@ static struct mapped_device *alloc_dev(int minor)
> >  		goto bad;
> >  
> >  	bio_init(&md->flush_bio, NULL, 0);
> > -	bio_set_dev(&md->flush_bio, md->bdev);
> >  	md->flush_bio.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH | REQ_SYNC;
> >  
> >  	dm_stats_init(&md->stats);
> 
> The top-level DM device's bdev->bd_disk is only assigned after the first
> blkdev_get(), so I can see why this is needed.
> 
> Think this type of life-cycle quirk was the kind of thing that caused
> blk cgroup issues with DM.
> 
> Left wondering whether there would be a better way of handling N flush
> bios for DM devices.  But for the now this will have to do.
> 
> Acked-by: Mike Snitzer <snitzer@xxxxxxxxxx>
> 

Thanks,
Dennis