Re: [Qemu-devel] [PATCH] qcow2: Support BDRV_REQ_MAY_UNMAP

[Fam Zheng <famz@redhat.com>]

On Wed, 09/28 18:11, Max Reitz wrote:
> On 28.09.2016 09:04, Fam Zheng wrote:
> > Handling this is similar to what is done to the L2 entry in the case of
> > compressed clusters.
> > 
> > Signed-off-by: Fam Zheng <famz@redhat.com>
> > ---
> >  block/qcow2-cluster.c | 9 +++++----
> >  block/qcow2.c         | 3 ++-
> >  block/qcow2.h         | 3 ++-
> >  3 files changed, 9 insertions(+), 6 deletions(-)
> > 
> > diff --git a/block/qcow2-cluster.c b/block/qcow2-cluster.c
> > index 61d1ffd..928c1e2 100644
> > --- a/block/qcow2-cluster.c
> > +++ b/block/qcow2-cluster.c
> > @@ -1558,7 +1558,7 @@ fail:
> >   * clusters.
> >   */
> >  static int zero_single_l2(BlockDriverState *bs, uint64_t offset,
> > -                          uint64_t nb_clusters)
> > +                          uint64_t nb_clusters, int flags)
> >  {
> >      BDRVQcow2State *s = bs->opaque;
> >      uint64_t *l2_table;
> > @@ -1582,7 +1582,7 @@ static int zero_single_l2(BlockDriverState *bs, uint64_t offset,
> >  
> >          /* Update L2 entries */
> >          qcow2_cache_entry_mark_dirty(bs, s->l2_table_cache, l2_table);
> > -        if (old_offset & QCOW_OFLAG_COMPRESSED) {
> > +        if (old_offset & QCOW_OFLAG_COMPRESSED || flags & BDRV_REQ_MAY_UNMAP) {
> >              l2_table[l2_index + i] = cpu_to_be64(QCOW_OFLAG_ZERO);
> >              qcow2_free_any_clusters(bs, old_offset, 1, QCOW2_DISCARD_REQUEST);
> 
> I don't quite understand the reasoning behind this. How is this more
> efficient than just using the existing path where we don't discard anything?

It's more efficient in disk space. I didn't mention because it is so not
specific to this, but: what virt-sparsify does is creating an overlay -> fstrim
it -> qemu-img commit. This flow revealed to me that BDRV_REQ_MAY_UNMAP should
have been honored this way (after a hint of "how" by Kevin).

> 
> Note that BDRV_REQ_MAY_UNMAP does not mean "Yes, please discard" but
> just "You may discard if it's easier for you". But it's actually not
> easier for us, so I don't see why we're doing it.
> 
> As far as I can guess you actually want some way to tell a block driver
> to actually make an effort to discard clusters as long they then read
> back as zero (which is why you cannot simply use bdrv_pdiscard()).
> However, I think this would require a new flag called
> BDRV_REQ_SHOULD_UNMAP (which should imply BDRV_REQ_MAY_UNMAP).

This flag doesn't make sense to me, if the protocol doesn't know how to unmap,
it can ignore BDRV_REQ_MAY_UNMAP, but not BDRV_REQ_SHOULD_UNMAP. It just
complicates things a little.

Fam