On 13.08.19 17:03, Kevin Wolf wrote:
> Am 13.08.2019 um 16:53 hat Max Reitz geschrieben:
>> On 13.08.19 16:46, Vladimir Sementsov-Ogievskiy wrote:
>>> 13.08.2019 17:31, Max Reitz wrote:
>>>> On 13.08.19 13:51, Kevin Wolf wrote:
>>>>
>>>> [...]
>>>>
>>>>> Hm... This is a mess. :-)
>>>>
>>>> Just out of curiosity: Why?
>>>>
>>>> Aren’t there only two things we really need from the block_status
>>>> infrastructure?
>>>>
>>>> (1) Whether something is allocated in the given layer of the backing chain,
>>>>
>>>> (2) Whether we know that a given range reads as zeroes.
>>>>
>>>> Do we really need anything else?
>>>>
>>>
>>> qemu-img map?
>>
>> Which is a debugging tool.  So it doesn’t fall under “really” in my
>> book.  If removing everything but allocation+zero information would make
>> the code a lot simpler, I think that would be worth it.
>>
>>> 1. We need to fix the bug somehow
>>> 2. We need to fix comment about different block-status flags, as it really
>>> lacks information of what actually "DATA" means (together with *file).
>>> And what finally means "allocated", can you define it precisely?
>>
>> As I wrote in my other mails, I think the problem is that it’s just
>> unexpected that block_status automatically skips through for filters.
>> It shouldn’t, that’s just black magic that the caller should not rely on.
>>
>> (We see precisely here that it’s wrong, because the callers are not
>> prepared for the allocation information returned to be associated with a
>> different node than what they passed.)
>>
>> So my definition is just “If the node has a COW backing file and
>> block_status returns ‘not allocated’, the data will be there.
>> Otherwise, the data is in the current node.”  Yes, that means that
>> filters should appear as fully allocated.
> 
> You can do that, but then the callers need to learn to do the recursion
> instead. After all, just copying everything if a filter is in the
> subtree isn't the desired behaviour.

Yes, hence the “deal with filters” series.

Max