Re: [Qemu-devel] [PATCH v2 5/5] vmdk: add bdrv_co_write_zeroes

[Stefan Hajnoczi <stefanha@gmail.com>]

On Mon, Apr 22, 2013 at 10:07:59AM +0800, Fam Zheng wrote:
> Use special offset to write zeroes efficiently, when zeroed-grain GTE is
> available. If zero-write an allocated cluster, cluster is leaked because
> its offset pointer is overwritten by "0x1".
> 
> Signed-off-by: Fam Zheng <famz@redhat.com>
> ---
>  block/vmdk.c | 82 ++++++++++++++++++++++++++++++++++++++++++++++--------------
>  1 file changed, 63 insertions(+), 19 deletions(-)
> 
> diff --git a/block/vmdk.c b/block/vmdk.c
> index 632689b..7475090 100644
> --- a/block/vmdk.c
> +++ b/block/vmdk.c
> @@ -814,6 +814,7 @@ static int get_whole_cluster(BlockDriverState *bs,
>  static int vmdk_L2update(VmdkExtent *extent, VmdkMetaData *m_data)
>  {
>      /* update L2 table */
> +    m_data->l2_offset = extent->l1_table[m_data->l1_index];

This is weird.  vmdk_L2update() should not modify m_data.  Please avoid
side-effects like this.

Is the VmdkMetaData.l2_offset field even needed?  Perhaps you can drop
it from the struct and use a local in vmdk_L2update().

>      if (bdrv_pwrite_sync(
>                  extent->file,
>                  ((int64_t)m_data->l2_offset * 512)
> @@ -905,6 +906,12 @@ static int get_cluster_offset(BlockDriverState *bs,
>      l2_index = ((offset >> 9) / extent->cluster_sectors) % extent->l2_size;
>      *cluster_offset = le32_to_cpu(l2_table[l2_index]);
>  
> +    if (m_data) {
> +        m_data->valid = 1;
> +        m_data->l1_index = l1_index;
> +        m_data->l2_index = l2_index;
> +        m_data->offset = cpu_to_le32(*cluster_offset);

It's risky to mix host endian and le32 fields in the same struct - the
chance is high that someone will make a mistake which field needs to be
byteswapped.

Instead, put the cpu_to_le32() inside vmdk_L2update() and use a local
variable there.  Please do this in a separate patch so it's easy to
review in isolation.

> +    }
>      if (extent->has_zero_grain && *cluster_offset == VMDK_GTE_ZEROED) {
>          zeroed = true;
>      }
> @@ -939,10 +946,6 @@ static int get_cluster_offset(BlockDriverState *bs,
>  
>          if (m_data) {
>              m_data->offset = tmp;
> -            m_data->l1_index = l1_index;
> -            m_data->l2_index = l2_index;
> -            m_data->l2_offset = l2_offset;
> -            m_data->valid = 1;
>          }
>      }
>      *cluster_offset <<= 9;
> @@ -1165,8 +1168,16 @@ static coroutine_fn int vmdk_co_read(BlockDriverState *bs, int64_t sector_num,
>      return ret;
>  }
>  
> +/**
> + * params:
> + *  - zeroed:       buf is ignored (data is zero), use zeroed_grain GTE
> + *                  feature if possible, otherwise return -ENOTSUP.
> + *  - zero_dry_run: used for zeroed == true only, don't update L2 table, just
> + *                  try if it's supported
> + */

Thanks for documenting the function!  Please use gtkdoc style, for
example:

/**
 * object_get_class:
 * @obj: A derivative of #Object
 *
 * Returns: The #ObjectClass of the type associated with @obj.
 */
ObjectClass *object_get_class(Object *obj);

(from include/qom/object.h)

https://developer.gnome.org/gtk-doc-manual/unstable/documenting_syntax.html.en

Previously there was no standard but it helps if we all use the same doc
comment style.

> @@ -1290,7 +1333,7 @@ static int vmdk_create_extent(const char *filename, int64_t filesize,
>      header.version = zeroed_grain ? 2 : 1;
>      header.flags = VMDK4_FLAG_RGD | VMDK4_FLAG_NL_DETECT
>                     | (compress ? VMDK4_FLAG_COMPRESS | VMDK4_FLAG_MARKER : 0)
> -                   | (zeroed_grain ? VMDK4_FLAG_ZG : 0);
> +                   | (zeroed_grain ? VMDK4_FLAG_ZERO_GRAIN : 0);

Please fix the patch series so the old VMDK4_FLAG_ZG is never
introduced.  The best way of ensuring this is to buildtest the series at
each patch:

git rebase -i master -x make

This will run make on every commit from master..HEAD.  If the build
fails, it stops and lets you git commit --amend.

Stefan