On 30.11.2015 10:01, Kevin Wolf wrote:
> Am 27.11.2015 um 18:58 hat Max Reitz geschrieben:
>> On 23.11.2015 16:59, Kevin Wolf wrote:
>>> Some drivers have nested options (e.g. blkdebug rule arrays), which
>>> don't belong to a child node and shouldn't be removed. Don't remove all
>>> options with "." in their name, but check for the complete prefixes of
>>> actually existing child nodes.
>>>
>>> Signed-off-by: Kevin Wolf <kwolf@redhat.com>
>>> ---
>>>  block.c                   | 19 +++++++++++++++----
>>>  include/block/block_int.h |  1 +
>>>  2 files changed, 16 insertions(+), 4 deletions(-)
>>
>> Thanks, now I don't need to fix it myself. :-)
>>
>> (I would have had to do that for an in-work series of mine)
>>
>>> diff --git a/block.c b/block.c
>>> index 23d9e10..02125e2 100644
>>> --- a/block.c
>>> +++ b/block.c
>>> @@ -1101,11 +1101,13 @@ static int bdrv_fill_options(QDict **options, const char **pfilename,
>>>  
>>>  static BdrvChild *bdrv_attach_child(BlockDriverState *parent_bs,
>>>                                      BlockDriverState *child_bs,
>>> +                                    const char *child_name,
>>>                                      const BdrvChildRole *child_role)
>>>  {
>>>      BdrvChild *child = g_new(BdrvChild, 1);
>>>      *child = (BdrvChild) {
>>>          .bs     = child_bs,
>>> +        .name   = child_name,
>>>          .role   = child_role,
>>>      };
>>>  
>>> @@ -1165,7 +1167,7 @@ void bdrv_set_backing_hd(BlockDriverState *bs, BlockDriverState *backing_hd)
>>>          bs->backing = NULL;
>>>          goto out;
>>>      }
>>> -    bs->backing = bdrv_attach_child(bs, backing_hd, &child_backing);
>>> +    bs->backing = bdrv_attach_child(bs, backing_hd, "backing", &child_backing);
>>>      bs->open_flags &= ~BDRV_O_NO_BACKING;
>>>      pstrcpy(bs->backing_file, sizeof(bs->backing_file), backing_hd->filename);
>>>      pstrcpy(bs->backing_format, sizeof(bs->backing_format),
>>> @@ -1322,7 +1324,7 @@ BdrvChild *bdrv_open_child(const char *filename,
>>>          goto done;
>>>      }
>>>  
>>> -    c = bdrv_attach_child(parent, bs, child_role);
>>> +    c = bdrv_attach_child(parent, bs, bdref_key, child_role);
>>>  
>>>  done:
>>>      qdict_del(options, bdref_key);
>>> @@ -3952,13 +3954,22 @@ static bool append_open_options(QDict *d, BlockDriverState *bs)
>>>  {
>>>      const QDictEntry *entry;
>>>      QemuOptDesc *desc;
>>> +    BdrvChild *child;
>>>      bool found_any = false;
>>> +    const char *p;
>>>  
>>>      for (entry = qdict_first(bs->options); entry;
>>>           entry = qdict_next(bs->options, entry))
>>>      {
>>> -        /* Only take options for this level */
>>> -        if (strchr(qdict_entry_key(entry), '.')) {
>>> +        /* Exclude options for children */
>>> +        QLIST_FOREACH(child, &bs->children, next) {
>>> +            if (strstart(qdict_entry_key(entry), child->name, &p)
>>> +                && (!*p || *p == '.'))
>>> +            {
>>> +                break;
>>> +            }
>>> +        }
>>> +        if (child) {
>>>              continue;
>>>          }
>>>  
>>
>> A good general solution, but I think bdrv_refresh_filename() may be bad
>> enough to break with general solutions. ;-)
>>
>> bdrv_refresh_filename() only considers "file" and "backing" (actually,
>> it only supports "file" for now, I'm working on "backing", though). The
>> only drivers with other children are quorum, blkdebug, blkverify and
>> VMDK. The former three have their own implementation of
>> bdrv_refresh_filename(), so they don't use append_open_options() at all.
>> The latter, however, (VMDK) does not.
>>
>> This change to append_open_options results in the extent.%d options
>> simply being omitted altogether because bdrv_refresh_filename() does not
>> fetch them. Before, they were included in the VMDK BDS's options, which
>> is not ideal but works more or less.
> 
> Are you sure? As far as I can tell, this patch should only keep options
> that were previously removed, but not remove options that were
> previously kept (with the exception of direct use of child names, which
> I added here to address your review comments for v1).
> 
> Specifically for "extents.%d", this is a child name and is therefore
> omitted. However, it contains a '.', so it was already removed without
> this patch.

Right, it is broken already. The same applies to qcow2's options
containing a dot.

Max

> I'm accepting proof of the contrary in the form of a test case. ;-)
> 
>> In order to "fix" this, I see three ways right now:
>> 1. Just care about "file" and "backing". bdrv_refresh_filename()
>>    doesn't support anything else, so that will be fine.
>> 2. Implement bdrv_refresh_filename() specifically for VMDK so
>>    append_open_options() will never have to handle anything but "file"
>>    and "backing".
>> 3. Fix bdrv_refresh_filename() so that it handles all children and not
>>    just "file" and "backing".
>>
>> Since we are shooting for 2.6 anyway (I assume ;-)), I think we should
>> go for option 3. This means that this patch is fine, and I'll see to
>> fixing bdrv_refresh_filename() (because I'm working on that anyway).
> 
> Yes, I agree.
> 
> Kevin
>