Re: [Qemu-devel] [PATCH 1/4] block: add image fragmentation statistics to qemu-img

[Kevin Wolf]

Am 12.03.2012 14:07, schrieb Stefan Hajnoczi:
> On Wed, Mar 7, 2012 at 9:22 AM, Dong Xu Wang <address@hidden> wrote:
>> From: Dong Xu Wang <address@hidden>
>>
>> Discussion can be found at:
>> http://patchwork.ozlabs.org/patch/128730/
>>
>> This patch add image fragmentation statistics while using qemu-img info.
>>
>> Signed-off-by: Dong Xu Wang <address@hidden>
>> ---
>>  block.c     |   13 +++++++++++++
>>  block.h     |    7 +++++++
>>  block_int.h |    1 +
>>  qemu-img.c  |    9 +++++++++
>>  4 files changed, 30 insertions(+), 0 deletions(-)
>>
>> diff --git a/block.c b/block.c
>> index 52ffe14..947607b 100644
>> --- a/block.c
>> +++ b/block.c
>> @@ -2588,6 +2588,19 @@ int bdrv_get_info(BlockDriverState *bs, 
>> BlockDriverInfo *bdi)
>>     return drv->bdrv_get_info(bs, bdi);
>>  }
>>
>> +int bdrv_get_fragment(BlockDriverState *bs, BlockFragInfo *bfi)
> 
> bdrv_get_fraginfo() makes it clearer that this function gets a summary
> of fragmentation information.  bdrv_get_fragment() makes me think it
> gets one specific "fragment".
> 
>> +{
>> +    BlockDriver *drv = bs->drv;
>> +    if (!drv) {
>> +        return -ENOMEDIUM;
>> +    }
>> +    if (!drv->bdrv_get_fragment) {
>> +        return -ENOTSUP;
>> +    }
>> +    memset(bfi, 0, sizeof(*bfi));
>> +    return drv->bdrv_get_fragment(bs, bfi);
> 
> For now this .drv_get_fraginfo() interface makes sense but if we merge
> the QCOW2<->QED in-place conversion patch series in the future it will
> be possible to implement this in a generic way because image formats
> will expose their guest -> host mapping.

We can probably resurrect Devin's patches for this part even now instead
of introducing an interface that we don't really want.

> 
>> @@ -1126,6 +1127,14 @@ static int img_info(int argc, char **argv)
>>             printf("cluster_size: %d\n", bdi.cluster_size);
>>         }
>>     }
>> +    if (bdrv_get_fragment(bs, &bfi) >= 0) {
> 
> I think we need a separate sub-command for fragmentation info:
> 
> qemu-img fraginfo <image-file>
> 
> Utilities that invoke qemu-img info want it to be fast.  Reading all
> metadata from a large image can take several seconds.  Since many
> qemu-img info users don't need to see the fragmentation information,
> it makes sense to put it in a new sub-command.

Yes. If we wanted to merge it into an existing qemu-img subcommand, I
think check would be the one, as it scans the whole image already today
and fragmentation is something that could be added fairly easily.

Kevin