Hello!

I think you are responsible for the change.

Yes.

even if tv1<tv2 which can be true when time wraps (which is true
at least with PSCHED_TIME_SOURCE == JIFFIES)

No, psched_time_t cannot wrap for our life.

one HTB user switched to 2.4.19-pre7.

I think this discipline simply calls the function with tv1 < tv2.
In this case the function is supposed to return a negative number
and used to do this. Actually, invalid behaviour with unsigned bound
was one of bugs fixed.

Here the return value will be always positive and < bound.

It must not be positive when tv1 < tv2, psched_diff_t is signed.
Well, taking into account that it is not used in current kernel
with tv1 < tv2, it can be changed to return 0, but surely not to "bound".

I did not see HTB qdisc, but plain logic tells me that otherwise
you will see a delay ~bound, when an immediate send is expected instead.

Alexey