Re: [Qemu-devel] [Qemu-block] [PATCH] block/mirror: enable detect zeroes when driving mirror

[wyang <w90p710@gmail.com>]

On 2016年11月22日 18:09, Kevin Wolf wrote:
> Am 22.11.2016 um 04:45 hat wyang geschrieben:
>> On 2016年11月21日 19:34, Kevin Wolf wrote:
>>> Am 21.11.2016 um 02:24 hat Yang Wei geschrieben:
>>>> In order to preserve sparse disk image, detect_zeroes
>>>> should also be enabled when bdrv_get_block_status_above()
>>>> returns BDRV_BLOCK_DATA
>>>>
>>>> Signed-off-by: Yang Wei <w90p710@gmail.com>
>>> Just preserving sparseness is exactly why bdrv_get_block_status_above()
>>> is checked, and this happens always.
>> I have a sparse qcow2 image(virtual size: 20G disk size: 12G),
>> # qemu-img info 20G.img
>> image: 20G.img
>> file format: qcow2
>> virtual size: 20G (21474836480 bytes)
>> disk size: 12G
>> cluster_size: 65536
>> Format specific information:
>>      compat: 1.1
>>      lazy refcounts: false
>>      refcount bits: 16
>>      corrupt: false
>>
>> and  I tried to qemu-img map --output json xxxx.img, as you can see,
>> the field of data always is true.
> Did you use metadata preallocation? If so, you explicitly requested all
> clusters to be allocated.

Yes,  I created the image with metadata preallocation. I see. its why 
qcow2_get_cluster_type always return QCOW2_CLUSTER_NORMAL.

>
>>> detect_zeroes does not preserve sparseness, but detect zeroes in parts
>>> of the source image that are not sparse, and make them sparse in the
>>> target image. This means that it has higher overhead (because all
>>> sectors need to be processed before copying them) and doesn't result in
>>> an exact copy. There may be cases where this is wanted, but in the
>>> common case, it's probably not the right mode of operation.
>> To be frank, customers are more concerned about sparseness than overhead.
> All generalisations are wrong.
>
> We could be talking about the majority of users (the QEMU project
> doesn't have customers),

The majority of users.:-) forgive my word.
>   but even then I doubt you're right, especially
> when using preallocation which is a tool for giving up sparseness in
> favour of improved performance of the first write to a cluster.

This patch is able to handle this case.  In fact I only hope to use 
preallocation to improve performance rather than become a sparseness 
image while live migrating.
it would waste a large mount of disk space, as most of our VM's data 
disk(/dev/vdb) is larger than 1TB. As you said,  We can use 
blockdev-mirror and manually enable detect-zeros
on new qemu verison(>2.6), but for less than 2.6, we also need to do it 
in drive-mirror.

>
>>> If you really want to detect zeroes, create the target block driver node
>>> manually (with detect-zeroes=on) and use the blockdev-mirror QMP command
>>> to mirror to it.
>> the target block is created by **bdrv_img_create** and opened by
>> **bdrv_open** inside qmp_drive_mirror, it also mean than we only
>> enable detect-zeroes while
>> its opened, i don't find qmp cmd to only enable it, if I understand
>> correctly.
> That's why I said you use blockdev-mirror instead drive-mirror.
Got it,  unfortunately  my qemu version does not contain this command. 
in order to use it, I also need to modify my libvirt.  its not 
temporarily available to do this. I would consider the my next release
to do it.:-)
>
> Kevin
>