Date:   Thu, 8 Jun 2023 20:28:16 -0400
From:   Mike Snitzer <snitzer@...hat.com>
To:     Sarthak Kukreti <sarthakkukreti@...omium.org>,
        Joe Thornber <ejt@...hat.com>,
        Brian Foster <bfoster@...hat.com>
Cc:     dm-devel@...hat.com, linux-block@...r.kernel.org,
        linux-ext4@...r.kernel.org, linux-kernel@...r.kernel.org,
        linux-fsdevel@...r.kernel.org, Jens Axboe <axboe@...nel.dk>,
        Theodore Ts'o <tytso@....edu>,
        "Michael S. Tsirkin" <mst@...hat.com>,
        "Darrick J. Wong" <djwong@...nel.org>,
        Jason Wang <jasowang@...hat.com>,
        Bart Van Assche <bvanassche@...gle.com>,
        Christoph Hellwig <hch@...radead.org>,
        Andreas Dilger <adilger.kernel@...ger.ca>,
        Stefan Hajnoczi <stefanha@...hat.com>,
        Alasdair Kergon <agk@...hat.com>
Subject: Re: [PATCH v7 4/5] dm-thin: Add REQ_OP_PROVISION support

On Thu, Jun 08 2023 at  5:24P -0400,
Mike Snitzer <snitzer@...nel.org> wrote:

> On Fri, May 19 2023 at 11:23P -0400,
> Mike Snitzer <snitzer@...nel.org> wrote:
> 
> > On Thu, May 18 2023 at  6:33P -0400,
> > Sarthak Kukreti <sarthakkukreti@...omium.org> wrote:
> > 
> > > dm-thinpool uses the provision request to provision
> > > blocks for a dm-thin device. dm-thinpool currently does not
> > > pass through REQ_OP_PROVISION to underlying devices.
> > > 
> > > For shared blocks, provision requests will break sharing and copy the
> > > contents of the entire block. Additionally, if 'skip_block_zeroing'
> > > is not set, dm-thin will opt to zero out the entire range as a part
> > > of provisioning.
> > > 
> > > Signed-off-by: Sarthak Kukreti <sarthakkukreti@...omium.org>
> > > ---
> > >  drivers/md/dm-thin.c | 74 +++++++++++++++++++++++++++++++++++++++++---
> > >  1 file changed, 70 insertions(+), 4 deletions(-)
> > > 
> > > diff --git a/drivers/md/dm-thin.c b/drivers/md/dm-thin.c
> > > index 2b13c949bd72..f1b68b558cf0 100644
> > > --- a/drivers/md/dm-thin.c
> > > +++ b/drivers/md/dm-thin.c
> > > @@ -1245,8 +1247,8 @@ static int io_overlaps_block(struct pool *pool, struct bio *bio)
> > >  
> > >  static int io_overwrites_block(struct pool *pool, struct bio *bio)
> > >  {
> > > -	return (bio_data_dir(bio) == WRITE) &&
> > > -		io_overlaps_block(pool, bio);
> > > +	return (bio_data_dir(bio) == WRITE) && io_overlaps_block(pool, bio) &&
> > > +	       bio_op(bio) != REQ_OP_PROVISION;
> > >  }
> > >  
> > >  static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
> > > @@ -1394,6 +1396,9 @@ static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
> > >  	m->data_block = data_block;
> > >  	m->cell = cell;
> > >  
> > > +	if (bio && bio_op(bio) == REQ_OP_PROVISION)
> > > +		m->bio = bio;
> > > +
> > >  	/*
> > >  	 * If the whole block of data is being overwritten or we are not
> > >  	 * zeroing pre-existing data, we can issue the bio immediately.
> > 
> > This doesn't seem like the best way to address avoiding passdown of
> > provision bios (relying on process_prepared_mapping's implementation
> > that happens to do the right thing if m->bio set).  Doing so cascades
> > into relying on complete_overwrite_bio() happening to _not_ actually
> > being specific to "overwrite" bios.
> > 
> > I don't have a better suggestion yet but will look closer.  Just think
> > this needs to be formalized a bit more rather than it happening to
> > "just work".
> > 
> > Cc'ing Joe to see what he thinks too.  This is something we can clean
> > up with a follow-on patch though, so not a show-stopper for this
> > series.
> 
> I haven't circled back to look close enough at this but
> REQ_OP_PROVISION bios _are_ being passed down to the thin-pool's
> underlying data device.
> 
> Brian Foster reported that if he issues a REQ_OP_PROVISION to a thin
> device after a snapshot (to break sharing), it'll fail with
> -EOPNOTSUPP (response is from bio being passed down to device that
> doesn't support it).  I was able to reproduce with:
> 
> # fallocate --offset 0 --length 1048576 /dev/test/thin
> # lvcreate -n snap --snapshot test/thin
> 
> # fallocate --offset 0 --length 1048576 /dev/test/thin
> fallocate: fallocate failed: Operation not supported
> 
> But reports success when retried:
> # fallocate --offset 0 --length 1048576 /dev/test/thin
> # echo $?
> 0
> 
> It's somewhat moot in that Joe will be reimplementing handling for
> REQ_OP_PROVISION _but_ in the meantime it'd be nice to have a version
> of this patch that doesn't error (due to passdown of REQ_OP_PROVISION)
> when breaking sharing.  Primarily so the XFS guys (Dave and Brian) can
> make progress.
> 
> I'll take a closer look tomorrow but figured I'd let you know.

This fixes the issue for me (causes process_prepared_mapping to end
the bio without REQ_OP_PROVISION passdown).

But like I said above back on May 19: needs cleanup to make it less of
a hack for the REQ_OP_PROVISION case. At a minimum
complete_overwrite_bio() would need renaming.

diff --git a/drivers/md/dm-thin.c b/drivers/md/dm-thin.c
index 43a6702f9efe..434a3b37af2f 100644
--- a/drivers/md/dm-thin.c
+++ b/drivers/md/dm-thin.c
@@ -1324,6 +1324,9 @@ static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
 	m->data_block = data_dest;
 	m->cell = cell;
 
+	if (bio_op(bio) == REQ_OP_PROVISION)
+		m->bio = bio;
+
 	/*
 	 * quiesce action + copy action + an extra reference held for the
 	 * duration of this function (we may need to inc later for a

Powered by blists - more mailing lists