Re: [PATCH] libxencall: Bump SONAME following new functionality

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

On 25/06/2021 11:59, Ian Jackson wrote:
> Jan Beulich writes ("Re: [PATCH] libxencall: Bump SONAME following new 
> functionality"):
>> On 25.06.2021 11:17, Andrew Cooper wrote:
>>> On 25/06/2021 07:31, Jan Beulich wrote:
>>>> On 24.06.2021 19:55, Andrew Cooper wrote:
>>>>> Fixes: bef64f2c00 ("libxencall: introduce variant of xencall2() returning 
>>>>> long")
>>>> Is this strictly necessary, i.e. is a Fixes: tag here warranted?
>>> Yes - very much so.
>>>
>>> andrewcoop@andrewcoop:/local/xen.git/xen$ readelf -Wa
>>> ../tools/libs/call/libxencall.so.1.2 | grep 1\\.3
>>>     33: 0000000000001496    59 FUNC    GLOBAL DEFAULT   13
>>> xencall2L@@VERS_1.3
>>>     39: 0000000000000000     0 OBJECT  GLOBAL DEFAULT  ABS VERS_1.3
>>>     76: 0000000000000000     0 OBJECT  GLOBAL DEFAULT  ABS VERS_1.3
>>>   020:   4 (VERS_1.2)      5 (VERS_1.3)      2 (VERS_1.0)      3
>>> (VERS_1.1)  
>>>   024:   3 (VERS_1.1)      2 (VERS_1.0)      4 (VERS_1.2)      5
>>> (VERS_1.3)  
>>>   0x0080: Rev: 1  Flags: none  Index: 5  Cnt: 2  Name: VERS_1.3
>>>
>>> Without this, you create a library called .so.1.2 with 1.3's ABI in.
>> I'm aware of the change to file contents as well as the disagreement
>> of file name / SONAME vs enumerated versions. So telling me this is
>> not really an answer to my question. It may be by convention that
>> the two should match up, but I don't see any functional issue (yet)
>> if they don't. Plus of course you leave open altogether the
>> backporting aspect of my question.
> The patch, including the Fixes tag,
>
> Reviewed-by: Ian Jackson <iwj@xxxxxxxxxxxxxx>

Thanks.

> Changing minor version in the filename as well as the .so is not an
> impediment to backporting.  The actual soname remains the same so
> there is no compatibility problem and the change is still suitable for
> including in eg distro stsable releases.

Correct, although backporting in general however is problematic.

Until Xen 4.16 is released (or, we explicitly decide to make an explicit
library release early), the 1.3 ABI isn't set in stone.

Backports to older stable-* branches must sit on a boundary already set
in stone in staging, or we'll end up with different versions of Xen
having different ideas of what VERS_1.3 mean.

It also creates upgrade compatibility problems created for any distro
who's installs of newer versions doesn't have internet access to pull
updated packages.  This is a consequence of having different versions of
stable libs bundled in different releases of Xen, rather than having
them entirely separate and packaged by their soname alone.

~Andrew