Re: [PATCH v8 1/3] docs: improve qcow2 spec about extending image header

[Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>]

18.10.2019 17:00, Eric Blake wrote:
> On 10/18/19 4:47 AM, Vladimir Sementsov-Ogievskiy wrote:
>> Make it more obvious how to add new fields to the version 3 header and
>> how to interpret them.
>>
>> The specification is adjusted so for new defined optional fields:
> 
> The specification is adjusted to make it clear that future fields are optional:
> 
>>
>> 1. Software may support some of these optional fields and ignore the
>>     others, which means that features may be backported to downstream
>>     Qemu independently.
>> 3. If @header_length is higher than the highest field end that software
>>     knows, it should assume that topmost unknown additional fields are
>>     correct, and keep additional unknown fields as is on rewriting the
>>     image.
>> 3. If we want to add incompatible field (or a field, for which some its
>>     values would be incompatible), it must be accompanied by
>>     incompatible feature bit.
>>
>> Also the concept of "default is zero" is clarified, as it's strange to
>> say that the value of the field is assumed to be zero for the software
>> version which don't know about the field at all and don't know how to
>> treat it be it zero or not.
>>
> 
> I'd also mention that we want to enforce 8-byte alignment in this cover letter.
> 
>> Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
>> ---
>>   docs/interop/qcow2.txt | 26 +++++++++++++++++++++++---
>>   1 file changed, 23 insertions(+), 3 deletions(-)
>>
>> diff --git a/docs/interop/qcow2.txt b/docs/interop/qcow2.txt
>> index af5711e533..4709f3bb30 100644
>> --- a/docs/interop/qcow2.txt
>> +++ b/docs/interop/qcow2.txt
>> @@ -79,9 +79,9 @@ The first cluster of a qcow2 image contains the file header:
>>                       Offset into the image file at which the snapshot table
>>                       starts. Must be aligned to a cluster boundary.
>> -If the version is 3 or higher, the header has the following additional fields.
>> -For version 2, the values are assumed to be zero, unless specified otherwise
>> -in the description of a field.
>> +For version 2, header is always 72 bytes length and finishes here.
>> +For version 3 or higher the header length is at least 104 bytes and has at
>> +least next five fields, up to the @header_length field.
> 
> For version 2, the header is exactly 72 bytes in length, and finishes here.
> For version 3 or higher, the header length is at least 104 bytes, including the next fields through header_length.
> 
>>            72 -  79:  incompatible_features
>>                       Bitmask of incompatible features. An implementation must
>> @@ -164,6 +164,26 @@ in the description of a field.
>>           100 - 103:  header_length
>>                       Length of the header structure in bytes. For version 2
>>                       images, the length is always assumed to be 72 bytes.
>> +                    For version 3 it's at least 104 bytes.
> 
> I'd also add a sentence that this field must be a multiple of 8.
> 
>> +
>> +Additional fields (version 3 and higher)
>> +
>> +The following fields of the header are optional: if software doesn't know how
>> +to interpret the field, it may be safely ignored, other than preserving the
>> +field unchanged when rewriting the image header.
> 
> Maybe:
> 
> if software doesn't know how to interpret the field, it may be safely ignored unless a corresponding incompatible feature flag bit is set; however, the field should be preserved unchanged when rewriting the image header.
> 
>> +
>> +For all additional fields zero value equals to absence of field (absence is
>> +when field.offset + field.size > @header_length). This implies
>> +that if software want's to set fields up to some field not aligned to multiply
>> +of 8 it must align header up by zeroes. And on the other hand, if software
>> +need some optional field which is absent it should assume that it's value is
>> +zero.
> 
> Maybe:
> 
> Each optional field that does not have a corresponding incompatible feature bit must support the value 0 that gives the same default behavior as when the optional field is omitted.

Hmmm. That doesn't work, as "corresponding" is something not actually defined. Consider our zstd extension.

It has corresponding incompatible bit, therefore, this sentence doesn't apply to it. But still, if incompatible bit is unset we can have this field. And it's zero value must correspond
to the absence of the field.

So, additional field may use incomaptible bit only for subset of its values.

But, I see, that you want to allow 0 value to not match field-absence if incompatible bit is set?

So, may be

Additional fields has the following compatible behavior by default:

1. If software doesn't know how to interpret the field, it may be safely ignored, other than preserving the field unchanged when rewriting the image header.
2. Zeroed additional field gives the same behavior as when this field is omitted.

This default behavior may be altered with help of incompatible feature bits. So, if, for example, additional field has corresponding incompatible feature bit, and it is set, we are sure that software which opens the image knows how to interpret the field, so,
1. The field definitely will not be ignored when corresponding incompatible bit is set.
2. The field may define any meaning it wants for zero value for the case when corresponding incompatible bit is set.

> 
>> +
>> +It's allowed for the header end to cut some field in the middle (in this case
>> +the field is considered as absent), but in this case the part of the field
>> +which is covered by @header_length must be zeroed.
>> +
>> +        < ... No additional fields in the header currently ... >
> 
> Do we even still need this if we require 8-byte alignment?  We'd never be able to cut a single field in the middle

hmm, for example:
105: compression byte
106-113: some other 8-bytes field, unalinged
113-119: padding to multiply of 8

- bad example, for sure. But to prevent it, we should also define some field alignment requirements..


>, but I suppose you are worried about cutting a 2-field 16-byte addition tied to a single feature in the middle.

and this too.

>  But that's not going to happen in practice.

why not?

4 bytes: feature 1

4 bytes: feature 2
8 bytes: feature 2

so, last 12 bytes may be considered as one field.. And software which don't know about feature2, will pad header to the middle of feature2

> The only time the header will be longer than 104 bytes is if at least one documented optional feature has been implemented/backported, and that feature will be implemented in its entirety.  If you backport a later feature but not the earlier, you're still going to set header_length to the boundary of the feature that you ARE backporting.

That's true, of course.

>  Thus, I argue that blindly setting header_length to 120 prior to the standard ever defining optional field(s) at 112-120 is premature, and that if we ever add a feature requiring bytes 112-128 for a new feature, you will never see a valid qcow2 file with a header length of 120.

consider my example above.



-- 
Best regards,
Vladimir