Am 16.09.2010 15:00, schrieb Anthony Liguori:
> On 09/16/2010 03:08 AM, Kevin Wolf wrote:
>> Am 11.09.2010 16:04, schrieb Anthony Liguori:
>>
>>> The use of protocols in backing_files is currently broken because of some
>>> checks for adjusting relative pathnames.
>>>
>>> Additionally, there's a spurious read when using an nbd protocol that can be
>>> quite destructive when using copy-on-read. Potentially, this can lead to
>>> probing an image file over top of NBD but this is completely wrong as NBD
>>> devices are not growable.
>>>
>>> Signed-off-by: Anthony Liguori<aligu...@us.ibm.com>
>>> ---
>>> NB: this is absolutely not ideal. A more elegant suggestion would be
>>> appreciated. I don't think NBD cleanly fits the model of a protocol as it
>>> stands today.
>>>
>>> diff --git a/block.c b/block.c
>>> index cd2ee31..a32d5dd 100644
>>> --- a/block.c
>>> +++ b/block.c
>>> @@ -344,6 +344,12 @@ static int find_image_format(const char *filename,
>>> BlockDriver **pdrv)
>>> return ret;
>>> }
>>>
>>> + if (strcmp(bs->drv->protocol_name, "nbd") == 0) {
>>> + drv = bs->drv;
>>> + bdrv_delete(bs);
>>> + goto out;
>>> + }
>>>
>> Is nbd really the only protocol that behaves like this? I don't like
>> hardcoding driver names in generic block layer code.
>>
>
> I'll drop this chunk from the patch as using backing_fmt achieves the
> same goal. The important hunk is the next one that removes assumptions
> that URIs are filenames.
Right, but next hunk is broken as Juan already mentioned. You always get
a protocol back, so with this patch applied you never resolve relative
paths for backing files.
Kevin
