On Fri, Jul 23, 2021 at 12:54:51AM +0300, Serge Semin wrote:

I always thought that ABI is supposed to be something what is
thoroughly documented and firmly declared to be so. It isn't something
claimed to be on a random nature but defined to be one when it's
more-or-less standardized. Thus the Linux kernel developers decide not
to change something unless it went through the series of iterations like
testing, stable, obsolete, remove. As I see it the rule-of-thumb is
supposed to be as "nothing is ABI unless it's declared as such".

Not true at all.  Again, if something works in an older kernel version,
and you upgrade to a new kernel version and it breaks, that is a
regression and must be fixed/reverted.

Lack of documentation does not mean an ABI can be changed.

greg k-h