Re: [Qemu-block] [PATCH] qcow2: add allocated-size to image specific info

[Denis V. Lunev]

On 05/18/2017 05:54 PM, Kevin Wolf wrote:
> Am 18.05.2017 um 14:22 hat Denis V. Lunev geschrieben:
>> On 05/18/2017 03:10 PM, Kevin Wolf wrote:
>>> Am 18.05.2017 um 13:04 hat Vladimir Sementsov-Ogievskiy geschrieben:
>>>> 18.05.2017 13:25, Kevin Wolf wrote:
>>>>> Am 18.05.2017 um 12:09 hat Vladimir Sementsov-Ogievskiy geschrieben:
>>>>>> Shows, how much data qcow2 allocates in underlying file. This should
>>>>>> be helpful on non-sparse file systems, when qemu-img info "disk size"
>>>>>> doesn't provide this information.
>>>>>>
>>>>>> Signed-off-by: Vladimir Sementsov-Ogievskiy <address@hidden>
>>>>>> ---
>>>>>> Hi all.
>>>>>>
>>>>>> Here is an allocated-size feature for qemu-img info.
>>>>> I'm not a fan of loading all L2 tables (can take some time) for
>>>>> 'qemu-img info' (which should be very quick). Why isn't the qemu-img
>>>>> check output good enough?
>>>>>
>>>>> Kevin
>>>>>
>>>>> $ ./qemu-img check /tmp/test.qcow2
>>>>> No errors were found on the image.
>>>>> 16164/491520 = 3.29% allocated, 11.98% fragmented, 0.00% compressed 
>>>>> clusters
>>>>> Image end offset: 1060044800
>>>>> $ ./qemu-img check --output=json /tmp/test.qcow2
>>>>> {
>>>>>     "image-end-offset": 1060044800,
>>>>>     "total-clusters": 491520,
>>>>>     "check-errors": 0,
>>>>>     "allocated-clusters": 16164,
>>>>>     "filename": "/tmp/test.qcow2",
>>>>>     "format": "qcow2",
>>>>>     "fragmented-clusters": 1937
>>>>> }
>>>> It is not the same, it shows guest clusters, but we need host
>>>> clusters - including all metadata, dirty bitmaps, snapshots, etc..
>>> Ah, right. But isn't that exactly the "disk size" (actual-size in JSON)
>>> from qemu-img info? Your commit message mentions non-sparse filesystems
>>> (which one?), but why wouldn't "disk size" provide this information
>>> there?
>>>
>>> The one case where it doesn't work is if you store a qcow2 image on a
>>> raw block device (this is something that oVirt does). In that case,
>>> you can't benefit from sparseness and disk space is used for a cluster
>>> in the middle even if its refcount is 0. oVirt uses "image-end-offset"
>>> to get the size of the first of the block device that is actually in use
>>> by the image.
>>>
>>> What is your exact use case? Maybe this helps me understand the exact
>>> kind of information that you need.
>>>
>>> Kevin
>> Let us assume we have an image like the following:
>>   [0][1][2][3][4][5][6]
>> Here [N] represents guest block number N, i.e. there are
>> 7 sequential guest blocks. Let us assume that the guest
>> issues TRIM and says that block [1] is not needed at all.
>> The image becomes like
>>   [0][.][2][3][4][5][6]
>> If the filesystem with this image is dumb and does not
>> support holes, we could not determine that we have
>> not used space inside the disk marked as [.]
>>
>> The goal of this patch is to know amount of [.] blocks.
> Okay. If the existing tools can't give you the numbers that you need,
> I think we could easily add this number to qemu-img check.
>
> However, is this really what you need or do you want to know the image
> size if the image was converted into a new image file? Because in this
> case, some metadata (refcount blocks) might go away as well. In this
> case, it sounds like Stefan's 'qemu-img measure' patches could be useful
> for you (I haven't looked at them in detail yet, but that's what I
> understood from the high level overview).
>
> Kevin
OK. 'qemu-img check' is not a problem to integrated with.
I'll look into measure patches to see applicability.
This is for sure good idea.

Thanks.