Received-SPF: pass (google.com: domain of linux-kernel+bounces-130027-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:4601:e00::3 as permitted sender) client-ip=2604:1380:4601:e00::3;
Date: Wed, 3 Apr 2024 11:03:46 -0400
From: Stefan Hajnoczi <stefanha@redhat.com>
To: Eric Blake <eblake@redhat.com>
Cc: linux-block@vger.kernel.org, linux-kernel@vger.kernel.org,
	Alasdair Kergon <agk@redhat.com>,
	Mikulas Patocka <mpatocka@redhat.com>, dm-devel@lists.linux.dev,
	David Teigland <teigland@redhat.com>,
	Mike Snitzer <snitzer@kernel.org>, Jens Axboe <axboe@kernel.dk>,
	Christoph Hellwig <hch@lst.de>, Joe Thornber <ejt@redhat.com>
Subject: Re: [RFC 8/9] dm thin: add llseek(SEEK_HOLE/SEEK_DATA) support
Message-ID: <20240403150346.GH2524049@fedora>
References: <20240328203910.2370087-1-stefanha@redhat.com>
 <20240328203910.2370087-9-stefanha@redhat.com>
 <c4pit5qf3sgiynx3jcnngdj7d3m62c5fdsgmla7twxynh6wfai@7jvhgxya4xo6>
Precedence: bulk
MIME-Version: 1.0
Content-Type: multipart/signed; micalg=pgp-sha256;
	protocol="application/pgp-signature"; boundary="RRfDDrHk+pzHnTwS"
Content-Disposition: inline
In-Reply-To: <c4pit5qf3sgiynx3jcnngdj7d3m62c5fdsgmla7twxynh6wfai@7jvhgxya4xo6>


--RRfDDrHk+pzHnTwS
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
Content-Transfer-Encoding: quoted-printable

On Thu, Mar 28, 2024 at 08:31:21PM -0500, Eric Blake wrote:
> On Thu, Mar 28, 2024 at 04:39:09PM -0400, Stefan Hajnoczi wrote:
> > Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
> > ---
> > Open issues:
> > - Locking?
> > - thin_seek_hole_data() does not run as a bio or request. This patch
> >   assumes dm_thin_find_mapped_range() synchronously performs I/O if
> >   metadata needs to be loaded from disk. Is that a valid assumption?
> > ---
> >  drivers/md/dm-thin.c | 77 ++++++++++++++++++++++++++++++++++++++++++++
> >  1 file changed, 77 insertions(+)
> >=20
> > diff --git a/drivers/md/dm-thin.c b/drivers/md/dm-thin.c
> > index 4793ad2aa1f7e..3c5dc4f0fe8a3 100644
> > --- a/drivers/md/dm-thin.c
> > +++ b/drivers/md/dm-thin.c
> > @@ -4501,6 +4501,82 @@ static void thin_io_hints(struct dm_target *ti, =
struct queue_limits *limits)
> >  	}
> >  }
> > =20
> > +static dm_block_t loff_to_block(struct pool *pool, loff_t offset)
> > +{
> > +	sector_t offset_sectors =3D offset >> SECTOR_SHIFT;
> > +	dm_block_t ret;
> > +
> > +	if (block_size_is_power_of_two(pool))
> > +		ret =3D offset_sectors >> pool->sectors_per_block_shift;
> > +	else {
> > +		ret =3D offset_sectors;
> > +		(void) sector_div(ret, pool->sectors_per_block);
> > +	}
> > +
> > +	return ret;
> > +}
> > +
> > +static loff_t block_to_loff(struct pool *pool, dm_block_t block)
> > +{
> > +	return block_to_sectors(pool, block) << SECTOR_SHIFT;
> > +}
> > +
> > +static loff_t thin_seek_hole_data(struct dm_target *ti, loff_t offset,
> > +		int whence)
> > +{
> > +	struct thin_c *tc =3D ti->private;
> > +	struct dm_thin_device *td =3D tc->td;
> > +	struct pool *pool =3D tc->pool;
> > +	dm_block_t begin;
> > +	dm_block_t end;
> > +	dm_block_t mapped_begin;
> > +	dm_block_t mapped_end;
> > +	dm_block_t pool_begin;
> > +	bool maybe_shared;
> > +	int ret;
> > +
> > +	/* TODO locking? */
> > +
> > +	if (block_size_is_power_of_two(pool))
> > +		end =3D ti->len >> pool->sectors_per_block_shift;
> > +	else {
> > +		end =3D ti->len;
> > +		(void) sector_div(end, pool->sectors_per_block);
> > +	}
> > +
> > +	offset -=3D ti->begin << SECTOR_SHIFT;
> > +
> > +	while (true) {
> > +		begin =3D loff_to_block(pool, offset);
> > +		ret =3D dm_thin_find_mapped_range(td, begin, end,
> > +						&mapped_begin, &mapped_end,
> > +						&pool_begin, &maybe_shared);
> > +		if (ret =3D=3D -ENODATA) {
> > +			if (whence =3D=3D SEEK_DATA)
> > +				return -ENXIO;
> > +			break;
> > +		} else if (ret < 0) {
> > +			/* TODO handle EWOULDBLOCK? */
> > +			return -ENXIO;
>=20
> This should probably be -EIO, not -ENXIO.

Yes. XFS also returns -EIO, so I guess it's okay to do so.

I still need to get to the bottom of whether calling
dm_thin_find_mapped_range() is sane here and what to do when/if it
returns EWOULDBLOCK.

> > +		}
> > +
> > +		/* SEEK_DATA finishes here... */
> > +		if (whence =3D=3D SEEK_DATA) {
> > +			if (mapped_begin !=3D begin)
> > +				offset =3D block_to_loff(pool, mapped_begin);
> > +			break;
> > +		}
> > +
> > +		/* ...while SEEK_HOLE may need to look further */
> > +		if (mapped_begin !=3D begin)
> > +			break; /* offset is in a hole */
> > +
> > +		offset =3D block_to_loff(pool, mapped_end);
> > +	}
> > +
> > +	return offset + (ti->begin << SECTOR_SHIFT);
>=20
> It's hard to follow, but I'm fairly certain that if whence =3D=3D
> SEEK_HOLE, you end up returning ti->begin + ti->len instead of -ENXIO
> if the range from begin to end is fully mapped; which is inconsistent
> with the semantics you have in 4/9 (although in 6/9 I argue that
> having all of the dm callbacks return ti->begin + ti->len instead of
> -ENXIO might make logic easier for iterating through consecutive ti,
> and then convert to -ENXIO only in the caller).

Returning (ti->begin + ti->len) << SECTOR_SHIFT for SEEK_HOLE when there
is data at the end of the target is intentional. This matches the
semantics of lseek().

I agree there is adjustment necessary in dm.c, but I want to seek the
semantics of all lseek() functions identical to avoid confusion.

Stefan

--RRfDDrHk+pzHnTwS
Content-Type: application/pgp-signature; name="signature.asc"

-----BEGIN PGP SIGNATURE-----

iQEzBAEBCAAdFiEEhpWov9P5fNqsNXdanKSrs4Grc8gFAmYNb9IACgkQnKSrs4Gr
c8iUVwf/Zpv3IsSznycexR7MwM8g1Pl1+FVxzDf79d7uGFplkj6L9AEhUrePfHRh
cGxgeFgNqzKAAfDc6rWkEoGxm8Jq5J/OwaTQnMfS5qhmCSekIanhihEtmh2o+3vQ
5ugyOOADTaWx4AvxKNrfr+16kosIWS8NOHreUzqvYxEsqJ8H07PwZBDNUslQociS
BtFimE8NmN40GYgH9KetDIyfzxwDE87WPWD61iQBKtsjpHaHcFcI1wsu+jqItN3i
u7kLKlLBFyLGnuQjuCCYsOQR629RC1ZE4z2j2IEzJEOeQvnAD2YfGjBk7fdt1mXw
bk76t2gtHwIgCVbdS5wnu6Bt+Wl6Ow==
=zZmQ
-----END PGP SIGNATURE-----

--RRfDDrHk+pzHnTwS--