Re: [PATCH v7 2/2] loop: Better discard support for block devices

[Evan Green <evgreen@chromium.org>]

On Wed, Nov 20, 2019 at 11:13 AM Darrick J. Wong
<darrick.wong@oracle.com> wrote:
>
> On Wed, Nov 20, 2019 at 10:56:30AM -0800, Evan Green wrote:
> > On Tue, Nov 19, 2019 at 6:25 PM Darrick J. Wong <darrick.wong@oracle.com> wrote:
> > >
> > > On Thu, Nov 14, 2019 at 03:50:08PM -0800, Evan Green wrote:
> > > > If the backing device for a loop device is itself a block device,
> > > > then mirror the "write zeroes" capabilities of the underlying
> > > > block device into the loop device. Copy this capability into both
> > > > max_write_zeroes_sectors and max_discard_sectors of the loop device.
> > > >
> > > > The reason for this is that REQ_OP_DISCARD on a loop device translates
> > > > into blkdev_issue_zeroout(), rather than blkdev_issue_discard(). This
> > > > presents a consistent interface for loop devices (that discarded data
> > > > is zeroed), regardless of the backing device type of the loop device.
> > > > There should be no behavior change for loop devices backed by regular
> > > > files.

(marking this spot for below)

> > > >
> > > > This change fixes blktest block/003, and removes an extraneous
> > > > error print in block/013 when testing on a loop device backed
> > > > by a block device that does not support discard.
> > > >
> > > > Signed-off-by: Evan Green <evgreen@chromium.org>
> > > > Reviewed-by: Gwendal Grignou <gwendal@chromium.org>
> > > > Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
> > > > ---
> > > >
> > > > Changes in v7:
> > > > - Rebase on top of Darrick's patch
> > > > - Tweak opening line of commit description (Darrick)
> > > >
> > > > Changes in v6: None
> > > > Changes in v5:
> > > > - Don't mirror discard if lo_encrypt_key_size is non-zero (Gwendal)
> > > >
> > > > Changes in v4:
> > > > - Mirror blkdev's write_zeroes into loopdev's discard_sectors.
> > > >
> > > > Changes in v3:
> > > > - Updated commit description
> > > >
> > > > Changes in v2: None
> > > >
> > > >  drivers/block/loop.c | 40 +++++++++++++++++++++++++++++-----------
> > > >  1 file changed, 29 insertions(+), 11 deletions(-)
> > > >
> > > > diff --git a/drivers/block/loop.c b/drivers/block/loop.c
> > > > index 6a9fe1f9fe84..e8f23e4b78f7 100644
> > > > --- a/drivers/block/loop.c
> > > > +++ b/drivers/block/loop.c
> > > > @@ -427,11 +427,12 @@ static int lo_fallocate(struct loop_device *lo, struct request *rq, loff_t pos,
> > > >        * information.
> > > >        */
> > > >       struct file *file = lo->lo_backing_file;
> > > > +     struct request_queue *q = lo->lo_queue;
> > > >       int ret;
> > > >
> > > >       mode |= FALLOC_FL_KEEP_SIZE;
> > > >
> > > > -     if ((!file->f_op->fallocate) || lo->lo_encrypt_key_size) {
> > > > +     if (!blk_queue_discard(q)) {
> > > >               ret = -EOPNOTSUPP;
> > > >               goto out;
> > > >       }
> > > > @@ -862,6 +863,21 @@ static void loop_config_discard(struct loop_device *lo)
> > > >       struct file *file = lo->lo_backing_file;
> > > >       struct inode *inode = file->f_mapping->host;
> > > >       struct request_queue *q = lo->lo_queue;
> > > > +     struct request_queue *backingq;
> > > > +
> > > > +     /*
> > > > +      * If the backing device is a block device, mirror its zeroing
> > > > +      * capability. REQ_OP_DISCARD translates to a zero-out even when backed
> > > > +      * by block devices to keep consistent behavior with file-backed loop
> > > > +      * devices.
> > > > +      */
> > > > +     if (S_ISBLK(inode->i_mode) && !lo->lo_encrypt_key_size) {
> > > > +             backingq = bdev_get_queue(inode->i_bdev);
> > > > +             blk_queue_max_discard_sectors(q,
> > > > +                     backingq->limits.max_write_zeroes_sectors);
> > >
> > > max_discard_sectors?
> >
> > I didn't plumb max_discard_sectors because for my scenario it never
> > ends up hitting the block device that way.
> >
> > The loop device either uses FL_ZERO_RANGE or FL_PUNCH_HOLE. When
> > backed by a block device, that ends up in blkdev_fallocate(), which
> > always translates both of those into blkdev_issue_zeroout(), not
> > blkdev_issue_discard(). So it's really the zeroing capabilities of the
> > block device that matters, even for loop discard operations. It seems
> > weird, but I think this is the right thing because it presents a
> > consistent interface to loop device users whether backed by a file
> > system file, or directly by a block device. That is, a previously
> > discarded range will read back as zeroes.
>
> Ah, right.  Could you add this paragraph as a comment explaining why
> we're setting max_discard_sectors from max_write_zeroes_sectors?

Sure. I put an explanation in the commit description (see spot I
marked above), but I agree a comment is probably also worthwhile.

>
> --D
>
> > -Evan